The wage effects of offshoring to the East and West: evidence from the German labor market

Abstract

This paper analyzes the labor market effects of offshoring in a high-wage home country and how these effects crucially depend on (1) Job complexity and (2) The characteristics of the destination country. It thereby links several sources: rich administrative data on individuals and plants in the German manufacturing industries, information on a job's task bundle, and the evolution of imported inputs from low- or high-wage destinations, which are represented by Eastern and Western Europe, respectively. Offshoring to these origins has opposing effects on German wages with respect to the relative task complexity of jobs: While offshoring to the West puts pressure on the wages of complex jobs and increases the wages of simple jobs, offshoring to the East entails the opposite effect. The overall effect adds up to a 4.2 percent increase in wages for jobs with high complexity, while low-complexity jobs see a 3.9 percent decrease in wages.

1 Introduction

1.1 Motivation and literature review

The most recent wave of globalization has broadly been driven by fostering international value chains and increasing trade in intermediate goods (Johnson & Noguera, 2017). Lower transport costs and new information technology have enabled industries to divide the manufacturing process into multiple parts, each of which fabricates a tradable output. Consequently, some production steps are offshored to benefit from international price differences. Specifically, relatively labor-intensive parts are moved to low-wage countries, whereas human-capital-intensive inputs are manufactured in high-wage countries (e.g., Carluccio et al., 2019). The resulting international value chain exploits comparative advantages through greater specialization in particular sets of tasks in source and destination countries. For the domestic labor market, this development emphasizes two counteracting forces. On the one hand, importing inputs substitutes for tasks formerly performed by domestic workers and thus places pressure on associated wages. On the other hand, it also reduces an industry's costs and boosts its productivity. Therefore, the industry's output expands, which, in turn, increases the demand for the remaining tasks in more specialized production and raises associated wages (Grossman & Rossi-Hansberg, 2008, 2012). In essence, any analysis of labor effects needs to consider the tasks substituted by imported inputs and the tasks that are allocated to complementary production.

Assuming that high-wage countries are skill abundant and specialize in particular human-capital-intensive goods, offshoring to these countries has different effects in terms of job substitution than offshoring to low-wage countries. Motivated by a steep increase in historically small trade flows (Fig. 1, or Krugman 2000), these effects have been the subject of fruitful discussion in recent decades. The literature has largely reached consensus that when not considering the characteristics of offshore production, offshoring lowers the relative demand for onshore workers without a college degree or for jobs with routine task profiles (e.g., Feenstra & Hanson, 1999; Becker et al., 2013; Baumgarten et al., 2013; Ebenstein et al., 2014; Hummels et al., 2014; Dauth et al., 2021).¹ Disagreement persists about the effects of offshored labor that is human-capital intensive, which is particularly surprising since the bulk of offshoring is between high-income countries and this type of trade has increased dramatically (Fig. 1). While Hummels et al. (2014, p. 1618 ff.) find a negative impact of offshoring to high-income countries on the Danish wages of low-skilled workers or routine jobs, Ebenstein et al. (2014, p. 588) reveal a positive wage impact on routine jobs in the US. Additionally, Mion & Zhu (2013) provide evidence from Belgian firms showing that imports from OECD countries negatively impact these firms' share of highly educated workers.

Fig. 1

Offshoring Intensity by Destination Region in German Manufacturing. Source: I-O Tables of the Federal Statistical Office of Germany (Fachserie 18, Reihe 2, Years: 1996–2007) and WIOT (2013), (Timmer et al., 2015). Notes: Offshoring intensity in German manufacturing is defined as the ratio of imported, intra-industry inputs relative to output. The left panel depicts region-specific offshoring from 1996 to 2007. The right panel displays the same shares less their 1996 values. From 1996 to 2002, offshoring to Central and Eastern Europe and offshoring to Western Europe increased by approximately 0.8 percentage points. As sudden access to CEECs also poses a supply shock from the perspective of (other) Western European countries, the expansion of offshoring to Western Europe constitutes a remarkable increase

These studies show that it is essential to distinguish the type of labor in onshore and offshore (the type/origin of imported inputs) production when estimating the heterogeneous impact of offshoring on wages. In earlier studies, onshore labor has been distinguished by a worker's education, whereas more contemporary works, such as Autor & Handel (2013), have shown that a job's task profile is more relevant when estimating wage compensation. Moreover, the task approach is used to distinguish labor by the costs of moving the job offshore, a characteristic that Blinder (2009) has named offshorability. In a more recent contribution, Blinder & Krueger (2013) find that well-paid workers and college graduates tend to hold jobs with higher offshorability and that they perform rather nonroutine tasks (e.g., mathematicians or programmers who can directly transfer their output via the Internet). While offshorable jobs are indeed prone to substitution with offshore labor (Goos et al., 2014), this vulnerability seems to be at odds with the fact that they are also the main gainers in terms of wages (e.g., Baumgarten et al., 2013). It is therefore doubtful whether the costs of moving a job offshore are the proper proxy for the manufacturing industry. In this sector, virtually every job is offshorable, as its tasks create a tangible good that can be sent to other regions (Blinder, 2006, p. 120). Then, the determining factor may again be the countries' comparative advantage in the production of goods that intensively require a specific set of tasks or type of labor. Regarding the wage effect of offshoring, these task inputs will then determine the substitutability of jobs.

1.2 Contribution and research question

The present paper makes important contributions to the existing literature in several regards. First, it adds to Baumgarten et al. (2013) and distinguishes offshoring with respect to the income level of its destination to approximate the human-capital intensity in imported inputs. New stylized facts show that these imports have crucially distinct effects on factor intensity in production. Complex-task intensive industries offshore to high-income countries and become less complex-task intensive over time, while the opposite is true for offshoring to low-income countries. Second, this paper combines existing complexity indices by Becker et al. (2013) and Brändle & Koch (2017), so that a single measure is able to distinguish groups of heterogeneous labor that respond differently to the substituting and complementary forces of typical inputs from high- or low-wage countries. Third, this paper sheds light on the underexplored topic on wage effects of offshoring to other high-income countries. Using an instrumental variable (IV) approach, this paper finds that offshoring to high-income countries has negative wage effects for complex jobs, while it positively affects wages for simple jobs.

The detailed analysis is feasible, because this study merges rich administrative data on workers in the West German manufacturing sector during the 1995-2007 period with plant-level information, micro-level data on tasks from the German Qualification and Career Survey (BIBB-IAB work survey), and offshoring data from federal input-output tables. To quantify the wage effects for very nuanced types of labor, I use an index of job complexity, which builds on data from the BIBB-IAB work survey and combines a wide variety of job information about the versatility of tasks, performance requirements (such as responsibility), and the required level of various skills and abilities (similar to Ottaviano et al., 2013). Across manufacturing jobs, the index is not intended to approximate the costs of moving a specific task set offshore; rather, it approximates the relative human-capital intensity (e.g., skill, knowledge, and abilities) imparted in production at a fine occupational level. The measure for offshoring intensity uses data from the German Federal Statistical Office, which—in contrast to UN Comtrade data or the World Input-Output Tables—directly record the industries' imports of inputs or purchases from a domestic supplier. Combining this source with the WIOT distinguishes offshoring destinations with respect to their income levels and approximates the complexity-intensity of imported inputs (see Table 1). Thereby, the paper focuses on Germany's most prominent destinations for vertical integration and groups them into economically relatively homogeneous units: the (human-)capital intensive European Union in the late 1990s (EU15) and the labor-intensive Central and Eastern European countries (CEECs).²

Table 1 Characteristics of Country Groups in the Year 2000

How does offshoring to these country groups impacts changes in the price of occupational task bundles? This question is answered by estimating Mincer-type wage equations, at which wages are determined at the industry or occupation level. Specifically, the large employer-employee dataset makes it possible to include worker-plant, occupation, and plant-year fixed effects to extract offshoring's wage impact within worker-plant matches while capturing endogenous plant-specific shocks (e.g., the exporter wage premium and new technology) of heterogeneous firms (e.g., Melitz, 2003) and asynchronous offshoring decisions within industries.

Despite the multidimensional fixed effects, wages and offshoring remain simultaneously determined, for example, because offshoring affects wages and wages affect the vulnerability to offshoring (e.g., offshoring activities could be more likely in relatively high-wage industries or occupations). This will bias the estimated coefficient of the causal wage impact of offshoring on wages. The analysis remedies these concerns by applying an IV regression, which extracts the exogenous variation in the offshoring variables. The choice of instruments builds on Autor et al. (2013), Baumgarten et al. (2013) and Hummels et al. (2014). It includes time-varying and region-specific instruments to suit the analysis with multiple trade partners. Accordingly, it utilizes the intermediate goods export supply of Germany's main offshoring destinations to other high-income countries. In the presence of the numerous fixed effects, these instruments depict an exogenous source of variation that is correlated with offshoring but independent of the wage-setting process in Germany.

The results confirm that offshoring has heterogeneous wage effects for manufacturing jobs that differ in complexity. Simple jobs benefit in terms of higher wage increases if domestic production expands the use of inputs from high-wage countries (EU15), while the relative wages of complex jobs suffer. Conversely, imported inputs from low-wage countries raise the wages of complex jobs but lower the wages of simple jobs. The overall effect adds up to a 4.2 percent increase in wages for a job with high complexity, while a low-complexity job sees a 3.9 percent decrease in wages.

1.3 Germany's economic integration in Europe and its labor market

Germany is a very suitable case to explore the wage effects of region-specific offshoring, particularly in the late 1990s and 2000s. First, the country is very representative because it is Europe's largest economy. Second, it ranks among the countries with the highest trade volumes worldwide and experienced a steep rise in offshoring intensity in the late 1990s and 2000s (less though in the 2010s, see Figure C1 in the appendix).³ Third, the fall of the Iron Curtain placed the country in a central position between an established trade bloc of high-wage, human-capital intensive countries in the west, the EU15, and low-wage, labor intensive countries in the east, that is, the CEECs (see, e.g., groupwise differences in the share of high-skilled workers in labor share in Table 1). Suddenly, Germany's geographic position became excellent to exploit international price differences within a short distance. This feature together with other political developments, that I describe in the following, vastly reduced the costs of offshoring and paved the way for the expansion of international value chains.

Eastward, the formerly separated CEECs featured relatively similar industrial and educational structures at substantially lower labor costs.⁴⁵ This phenomenon placed the German economy in a more competitive environment that was bolstered by several reductions in trade costs: In the early 1990s, the CEECs signed association agreements with the EU, which vastly cut tariffs. Trade flows, however, did not substantially increase until EU accession talks began in 1997. These negotiations endorsed the market system and institutions of the newly established democracies and, hence, gradually stabilized the investment climate. Moreover, it gave rise to the installation of foreign affiliates, even before these countries entered the EU in 2004. With these firms bringing in new production technology from their parent companies (Dustmann et al., 2014), the internal productivity and international competitiveness of suppliers in the CEECs rose steeply, resulting in vast expansions of imports from those regions to Germany.

Simultaneously, the EU politically reinforced the value chains among the EU15 countries. Beyond the already existing advantages of a customs union, the EU suppressed internal nontariff barriers by harmonizing regulations, laws, standards, and economic practices. European infrastructure projects and the establishment of the Schengen Area in 1995 lowered the costs of transportation, e.g., through new cross-border roads or time savings due to the abandonment of border controls. Furthermore, in 1999, the introduction of a common currency, the euro, abolished exchange rate fluctuations. Together, these measures vastly reduced the costs of offshoring.

How these events come along with offshoring from Germany to these destinations is depicted in Fig. 1. It clearly shows that offshoring to the EU15 exhibits substantially higher offshoring intensities than to any other country group. From 1996 to 2007, the share of inputs from the EU15 relative to industry output in Germany grew by 0.91 percentage points, or 25 percent of its initial value (Table C1 in the appendix). The right panel emphasizes the increasing relevance of CEECs as offshoring destinations. While this country group exhibits low initial values of economic integration with Germany, from 1996 to 2007, offshoring to these countries increased by 1.1 percentage points, or 318 percent.⁶

While the German goods market is characterized by large and growing trade volumes, the increase in output demand did not immediately translate into growth of the labor market. In fact, the labor market has instead been characterized by rather high rates of unemployment and wage polarization.⁷ It seems that the evolution of trade comes along with a change in the demand for (or the marginal product of) certain types of labor.

Fig. 2

Wage Divergence between Terciles of Job Complexity. Source: BIBB-IAB Work Survey, LIAB. Notes: Indexed wage growth of terciles of the task index, West Germany, manufacturing, 1996–2007, 85 percent sample

Figure 2 illustrates the divergence in average real wages for the terciles of the complexity distribution. It reveals that income growth is unequally distributed and varies by job complexity. While the wages of complex jobs rise by 13 percent, the compensation for intermediate jobs rises by approximately 8 percent, and wages of simple jobs increase by less than 5 percent.⁸

The remainder of this paper is structured as follows. Section 2 presents the various datasets employed in the analysis. Then, Sect. 3 explains the estimating equation and the identification strategy for the empirical analysis. Section 4 compiles the results, which are checked for robustness in Sect. 5. Finally, Sect. 6 concludes the paper.

2 Data description

This section introduces the various datasets employed in the analysis and provides summary information on data construction and measurement. For details on the sampling procedure and data processing, I refer to the appendix.

2.1 Linked employer-employee data

I extract matched information on workers and plants from a longitudinal version of the Linked Employer-Employee (LIAB MM 9308) dataset of the Institute for Employment Research (IAB).⁹ The LIAB has important features for the analysis at hand. First, it is designed to provide a long time dimension with many entries per employer, which is well suited to the objective of capturing unobserved heterogeneity in plants or individuals through multidimensional fixed effects.

Second, the LIAB samples the most comprehensive dataset of workers in Germany, comprising the universe of employees subject to social security (approximately 80 percent of the workforce). These data are drawn from social security registers and contain worker characteristics, such as age, sex, education, work experience, job tenure, occupation, occupational status (part-time, full-time, or apprentice workers), and average daily wages during an employment spell. As stating incorrect information incurs a penalty, the recorded wage data are very reliable. Above a contribution ceiling, however, wages are top-coded and need to be imputed.

Third, the LIAB contains administrative data on plants, such as the number of employees, the location, and the industry code. It is also possible to merge a subsample of the businesses with additional information from an annually conducted survey, the IAB Establishment Panel (EP).¹⁰ In comprehensive interviews, the plants' managers provide precise information about the composition of the plant's workforce, revenues, investments, export share, and type of union coverage.¹¹ Since I merge annual information on plants with worker data, which are available on a daily level, I restrict all observations to yearly intervals to arrive at a consistent time scale.

Finally, one particular advantage of the LIAB is that occupational codes are classified according to the similarity of tasks on the job (Bundesagentur für Arbeit, 1998). Since its scheme KldB88 is identical to the classification in the BIBB-IAB work survey, it is possible to assess a job's typical complexity akin to the procedure developed by Autor et al. (2003).

2.2 Job complexity index

The job complexity index is intended to measure the heterogeneity of labor in the wage regression. By focusing exclusively on manufacturing, in which virtually all jobs are offshorable, this concept is associated with the comparative advantage in the production of goods and the regional specialization in particular tasks.¹²

The combined index has several advantages. In addition to featuring relatively high correlations with existing indices such as those developed by Spitz-Oener (2006), Baumgarten et al. (2013), Becker et al. (2013), and Brändle & Koch (2017), it includes not only the intensity of one job characteristic (e.g., routineness) but also detailed information on a variety of tasks and requirements.¹³ By using a broad set of information, it is feasible to extract the variation of 243 occupations, which produces a decisive increase in the degrees of freedom in the subsequent analysis.

The data are drawn from the BIBB-IAB work surveys, which are jointly compiled by the Federal Institute for Vocational Education and Training and the IAB.¹⁴ Approximately every six years, randomly selected workers from the German labor force answer questions about their abilities, performance requirements, professional qualifications and tasks on the job.¹⁵ I utilize the two cross-sectional waves of 1998/99 and 2006 since they lie within the sample period and refer to the same population as the LIAB. Each wave covers 20,000 to 34,000 individuals.

The task complexity index applies a methodology similar to those of Becker et al. (2013) and Ottaviano et al. (2013). Therefore, the categories of interactive and nonroutine tasks are merged and extended with job performance requirements and necessary abilities. Table B2 in the appendix provides an overview of the various components.

In instancing interactive tasks or tasks that require many face-to-face interactions, cultural ties, or interpersonal skills, it is difficult to evaluate the tasks individually because, for example, collaboration with coworkers does not necessarily imply high complexity. Instead, various applications, such as dealing with consumer preferences, the legal system, various languages, and face-to-face interactions, may better approximate the human-capital requirements of a job.

Furthermore, performing many nonrepetitive tasks that require customized problem-solving ability is considered relatively complex. I assess nonroutine tasks based on whether young apprentices could perform these tasks independently in their first week of work. Since the survey questions about tasks are relatively broad, the nature of tasks may still vary substantially between occupations. The nonroutine task “consult or inform”, for instance, features high affirmative response rates by telephone operators, cashiers, auditors, and managers. Thus, the different kinds of seemingly identical tasks suggest a further distinction.¹⁶

To approximate quality, I consider information on the requirements for job performance and on certain skills or special/sensitive knowledge. Typically, the higher such requirements are, the more they foster regional specialization within (high-wage) countries due to local knowledge spillovers, locally concentrated experience in certain tasks, and external economies of scale (Grossman & Rossi-Hansberg, 2012).

Consequently, I combine the degree of interactivity and nonroutineness of the job and the level of required abilities into a single complexity measure. To do so, I assign the responses of each wave to four groups, which differ with respect to the scaling and style of the survey questions.

In the 1998/1999 wave, the first group consisted of polar questions about the use of 81 workplace tools. Such tools range from machinery and diagnostic devices to computers, communication equipment, means of transport, and software. Whenever a worker reports having used a tool that is associated with a rather complex activity, this entry is marked.

In the 2006 wave, the questions in the first group are directly intended to capture the scope of nonroutine and interactive tasks. For example, workers state how often they present something or how often they have to solve new or unforeseen problems. The questions in the other three groups are similar in the two waves.

In the second group, the questions are intended to explore the frequency of 13 specific activities on the job. These are described, for instance, as repairing, consulting, educating, analyzing, or producing. The more frequently a worker performs any complex activity, such as consulting and educating, the higher the respective value is.

The third group comprises questions on specific abilities or knowledge. This includes, for example, any job that requires profound knowledge of the German legal system or high levels of English, German or any other particular foreign language skill.

In the fourth group, workers answer questions about performance requirements on the job, e.g., the frequency of having deadlines, whether mistakes lead to vast financial losses for the firm, or whether they have to improve techniques or processes.

To establish the comparability of the BIBB-IAB work survey and the LIAB, I consider only employees with social insurance who work more than 20 hours per week. For each of the eight task groups, I mark all affirmations of complexity before I separately sum them up per individual. In a second step, I average such sums over each 3-digit occupation. A higher mean indicates an occupation that (1) Is more likely to entail a larger number of different complex tasks, (2) Spends a larger share of its working time on the performance of complex tasks, and/or (3) Requires higher knowledge, skills, or abilities for its task set. Subsequently, I drop all occupations that encompass fewer than five individuals and normalize the remaining values by dividing the occupation averages by the maximum value of all means. The outcome is an index that ranges between zero and one for each of the eight groups.

I then separately sum up the four indexes per wave and normalize over all occupations to receive a single index for each of the two survey years. In a final step, I take the frequency-weighted average of the indices from the two waves (using the observations per occupation as weights) and obtain one static index for the analysis. A high value of the index is associated with a high relative input of human capital. While the simple combination of existing indices is an arguably arbitrary construction, it has the advantage of considering many task dimensions that are related to offshorability. Note also that the resulting ranking of occupations is highly correlated with existing indices and, most importantly, the average educational attainment per occupation (Table C2 in the appendix). Table 2 presents a list of occupations in manufacturing with the highest or lowest values of the task complexity index.

Table 2 The ten most and least complex occupations in manufacturing

2.3 Offshoring measure

The analyses use industry- or occupation-level offshoring measures, which are constructed in several steps. The starting point is the “narrow” definition of offshoring for manufacturing industries, as in Feenstra & Hanson (1996, 1999). Its construction considers imported inputs M that are produced in the same classification of economic activity $j^{*}$ as the domestic using industry j (NACE/ISIC rev. 3). Hence, it contemplates firms' productivity decisions with respect to either producing those inputs or importing them. These industry's imports are then normalized by its gross output Y in year t:

$$\begin{aligned} OS_{jt} = \dfrac{M^{j^{*}}_{jt}}{Y_{jt}} . \end{aligned}$$

(1)

The offshoring measure $OS_{jt}$ indicates the share of value added abroad that could, instead, be produced by the respective domestic industry. The data are drawn from the input-output tables of the German Federal Statistical Office, which is crucial for the analysis because it explicitly distinguishes between domestically and foreign-produced inputs (Winkler & Milberg, 2012, p. 40).

To arrive at region specific measures, I map the countries into groups with respect to their affiliation with a trade bloc, their geographic proximity, and the similarity of their economic structures. Thereby, I focus only on the most relevant offshoring destinations for Germany during the period considered because including more regions impedes the separate identification of the respective causal effects. These regions include the countries belonging to the former EU15 and the CEECs.

The offshoring measure $OS_{jt}$ becomes region specific due to the weighting by the share of intra-industry imports from a particular region r in the total intra-industry imports of sector j in year t:

$$\begin{aligned} Ofs_{jtr} = \phi _{jtr} OS_{jt} , \end{aligned}$$

(2)

where

$$\begin{aligned} \phi _{jtr} = \dfrac{M^{j^{*}}_{jtr}}{M^{j^{*}}_{jt,World}} . \end{aligned}$$

(3)

I draw the region specific data from the WIOT.

For the occupational exposures to offshoring, I weight the region-specific offshoring exposures of 24 industries $Ofs_{jtr}$ by the number of workers in occupation q and industry j relative to the total number of employees in occupation q in manufacturing in 1995. This weighting yields 243 annual occupation-specific exposures to region-specific offshoring, which is insensitive to subsequent alterations in the composition of jobs across industries:

$$\begin{aligned} OccOfs_{qtr} = \sum \limits _{j=1}^{J}\frac{L_{qj,1995}}{L_{q,1995}} Ofs_{jtr} . \end{aligned}$$

(4)

In the Online Appendix B.3, I provide more technical details about the offshoring measure. Table C3 examines the summary statistics, and Table B1 provides more comprehensive details on the data processing.

Due to the consideration of intraindustry imports and occupational exposures, it is feasible to reveal how imparted tasks in intermediates impact the compensation of similar or dissimilar task sets in occupations.

3 Framework and empirical strategy

The theoretical literature identifies the various channels of offshoring that can affect labor demand (e.g., Groizard et al., 2014). In essence, the embodiment of these channels depends on a job's task profile. Although I adopt the idea of task trade as the trade of specific labor inputs, I do not attempt to disentangle the diverse channels at work; instead, similarly to Hummels et al. (2014), I utilize an estimable production function framework that describes the aggregate labor market effects of offshoring on different jobs.

3.1 Estimation equation

The analysis contemplates linkages of domestic production with international value chains. Importing inputs substitutes for only a fraction of workers in the respective producing industry, while it complements the remaining workers in this industry.¹⁷ Hence, it results in opposing effects within industries, where some occupations benefit and others lose. This reasoning indicates that industries are a well-suited unit for analyzing the ambivalent impacts of offshoring. However, to disentangle and quantify these counteracting effects in the home country, the exposure to imported inputs should be further partitioned into relevant task bundles. By including the occupation-weighted offshoring values from Eq. (4), it is straightforward to arrive at the occupational elasticities $\beta _r$ and $\beta _{R+r}$, which capture cross-industry spillovers:

$$\begin{aligned}&\ln w_{iqlt} = \sum _r \beta _r OccOfs_{qtr} + \sum _r \beta _{R+r} OccOfs_{qtr} \times CMPLX_q\nonumber \\&\qquad\quad\;\; + {\varvec{x}}_{it} \varvec{\lambda } + \gamma _{il} + \theta _q + \kappa _{lt} + \epsilon _{iqt}, \end{aligned}$$

(5)

where r indicates the region, $r \in \{CEECs, EU15 \}$, and $w_{iqlt}$ denotes the daily real wage of individual i in plant l, year t, and occupation q.¹⁸ This equation is the baseline estimation for the analysis. It comprises three dimensions of fixed effects that account for unobserved heterogeneity. First, worker-plant fixed effects $\gamma _{il}$ capture unobservable worker-plant-specific productivity. Second, occupation fixed effects $\theta _q$ incorporate observable and unobservable time-invariant characteristics of occupations. This term also absorbs the explanatory power of $CMPLX_q$, whereas it is still possible to identify the interaction term with offshoring due to the variation in occupational exposure. Third, plant-year fixed effects $\kappa _{lt}$ exhibit two important properties. On the one hand, they capture plant-specific shocks that are correlated with wages and offshoring, such as plant-specific technological change. On the other hand, they avoid endogeneity concerns (with wages) since plant controls are multicollinear.¹⁹

Before estimating Eq. (5), I eliminate the effects of wage changes due to variations in working hours or gender-specific wage developments and reduce the unbalanced sample to full-time male employees in West Germany (excluding West Berlin).²⁰ Furthermore, the panel data consider only the best-paid spells in full time (excluding confounding overlaps of employment) in manufacturing (industries 15 - 37) on June 30 of each year from 1995 to 2007.

The LIAB reports the wage information of individuals up to the social insurance contribution ceiling, which yields 11 to 15 percent of top-coded employment spells per year.²¹ Since the censored distribution biases the ordinary least squares (OLS) estimation, I either impute the missing values similarly to Card et al. (2013) (see the description in B.1.3) or cut off the sample at the 85th percentile of the wage distribution (85 percent sample). The latter renders attenuated estimates of the coefficients and variance (see Biørn, 2016) and therefore provides a lower bound of the absolute magnitude of the effects that need to be reaffirmed by other specifications due to the downward bias of the variance estimates.

3.2 Identification and instruments

Identifying the causal effect of offshoring on wages requires that endogeneity concerns be addressed. While offshoring affects wages, wages simultaneously affect the propensity to move associated tasks abroad, for example, if offshoring is more probable in high-wage industries or high-wage occupations due to higher potential cost savings.²² Accordingly, the OLS estimates would be biased towards zero, because complex-task intensive inputs from the EU15 are expected to put pressure on the wages of especially those complex jobs that are better paid than other complex jobs (analogous for simple jobs and offshoring to the CEECs). I solve the endogeneity problem using an IV strategy, where time-varying and region-specific instruments must correlate with region-specific offshoring while being independent of wage setting in Germany.²³ In the selection of suitable instruments for $OccOfs_{qtr}$. I follow the work by Hummels et al. (2014) and consider the export supply of intra-industry goods ESI from the respective offshoring destination r to other high-income economies HI:²⁴

$$\begin{aligned} ESI^{HI}_{qtr} = \sum \limits _{j=1}^{J}\frac{L_{qj,1995}}{L_{q,1995}} \dfrac{S^{HI}_{jtr}}{Y_{jtr}} , \text {\,\,for each region { \it{r}}.} \end{aligned}$$

(6)

Applying the weights from Eq. (4), $L_{qj,1995}$ denotes the number of employees in occupation q and industry j in 1995 relative to the total number of manufacturing workers in occupation q, $L_{q,1995}$, in Germany. $S^{HI}_{jtr}$ denotes the supply of intra-industry exports that are demanded by high-wage countries other than Germany, and $Y_{jtr}$ represents the output value of the respective foreign industry j. The trade data originate from the WIOT.²⁵

To test the instrument invalidity (or any other misspecification), I add an overidentifying instrument akin to Baumgarten et al. (2013), namely, the ad valorem trade costs of shipping containers that Europe imports from China. Although maritime trade does not capture the modes of transportation for most goods within Europe, the costs of shipping containers still seem to exhibit high explanatory power for the other European transport costs. These costs thus comprise an eligible instrument because they are correlated with offshoring while being orthogonal to German wages.²⁶ Note, however, that shocks to transportation expenses not only lower the costs of imported inputs in Germany but also decrease the costs of German goods abroad. This phenomenon positively affects foreign demand and, consequently, the outcome variables of German plants, such as the export share, revenue, investment, or number of employees. In the baseline regression, I avoid the endogeneity of plant controls by including plant-year fixed effects, which also control for any shock within an industry that is not visible at the aggregate level. Therefore, they also account for the timing of offshoring activities and the introduction of new technology, which affect, e.g., the plant's number of employees, revenue, capital per worker, and task composition. The corresponding identifying assumption is that technology shocks affect wages in plants (or industries) but not at the occupation level.

This assumption raises concerns about unobservable skill-biased technological change and generally regards instrument validity, as precisely discussed in Autor et al. (2013). In essence, a (technology) shock that is common to all high-income countries may affect the demand for inputs to a similar extent. Therefore, if technology is simultaneously available to all high-income countries and correlates with both occupational wages in Germany and occupational exposure to offshoring in other high-income countries (the instruments), the IV estimates could still be biased. Although it is impossible to completely disentangle those effects, various specifications substantially mitigate confounders. First, the inclusion of plant-year fixed effects alleviates a potential bias since it implicitly considers the yearly plant-specific composition of workers, tasks and technology, such as computer use rates. The identified wage effects deviate from the average annual development in the plant. For instance, the simultaneous introduction of new technology would be controlled for by capturing the annual wage bill. In the specification with plant controls, the interaction of capital per worker and job complexity captures the channel of new technology on occupational wages. Second, the occupational exposures are weighted by initial worker shares per industry. This weighting creates an extra layer that mitigates the bias from technology if the other high-income countries feature a different worker-industry structure (different weighting in 1995) and if technology does not affect individual wages in Germany parallel to the sourcing of respective inputs in other high-income countries. Note that the latter condition describes the violation of the instrument validity (correlation of instruments with wages beyond offshoring from Germany).²⁷

4 Results

This section begins by describing the links among industry characteristics prior to an increase in offshoring to either high- or low-wage countries and emphasizes the adjustments within plants that accompany such increases. The analysis then assesses the wage effect of industry or occupational exposure to offshoring.

4.1 Preliminary analysis: offshoring, industry characteristics, and plants' adjustments

Recalling the insights from Heckscher-Ohlin theory, low-wage countries, which are abundant in low-skilled labor, specialize and export simple task-intensive goods, whereas high-wage countries, where low-skilled labor is relatively scarce, export complex task-intensive goods. Furthermore, according to Grossman & Rossi-Hansberg (2012), the intra-industry trade between high-wage countries is explained by specialization in certain sets of tasks due to knowledge spillovers and scale economies. Table 3 shows how actual offshoring to two representatives of such regions correlates with plant-level outcomes in Germany. It provides insights on the initial characteristics of exposed industries and how these characteristics change when offshoring increases. Especially, columns 3–6 of panel B reveal changes in the quantitative dimension of relative labor supply, thus, the flip side of the estimates in the baseline regression in the next subsection.²⁸ Each cell represents a separate simple linear regression with firm-level outcomes as dependent variables regressed on industry-level offshoring. The correlations in columns 1 and 2 provide weak suggestive evidence on the initial industry characteristics that are followed by increases in offshoring. The main interest, however, is columns 5 and 6, which present the plant adjustments that come along with increases in offshoring.

Table 3 Simple Regressions of Offshoring on Plant-Level Outcomes in Germany

In columns 1 and 2 of panel A, I regress year-1995 industry averages of one plant outcome variable on industry-level offshoring in 2005 either to the CEECs or the EU15 (explanatory variable).²⁹³⁰ Employing future values of offshoring (that include the treatment) on a cross-section primarily yields a level effect or initial characteristics of industries that are exposed to region-specific offshoring (vs. how they are affected). Additionally, state fixed effects account for regional differences among the German federal states. The coefficients suggest that offshoring takes place mainly in industries where plants have many employees, higher export shares, revenues, and wage bills and are covered by collective bargaining at the industry level. Note that all these characteristics are also typical of large and more productive firms (Melitz, 2003).

Each cell in columns 3–6 shows the estimate from a panel regression from 1996 to 2007 that includes plant fixed effects in addition to the single regressor.³¹ The estimates present dependencies between changes in the outcome variables and changes in offshoring. Causality, however, is not inferred from the results. In contrast, the relationship may imply that the outcome variables determine offshoring, e.g., because plants with higher revenues can afford the costs of offshoring or offshoring determines the outcome variables, e.g., offshoring increases the revenues of plants. Alternatively, both could be true, and the variables would then be simultaneously determined due to reverse causality or due to any other shock to plants' demand or productivity and offshoring. These links are an integral part of the identification challenge and require consideration in the subsequent analysis. I mitigate this problem in columns 5 and 6 by predicting the values for the two types of offshoring using the instruments from Sect. 3.2.³²

The estimates suggest that the correlations vary substantially for predicted and non-predicted values of offshoring and between the two destination regions. Examining column 3, the intensity of offshoring to the EU15 (hereafter $Ofs^{EU15}_{jt}$) and plant outcomes reveal hardly any distinct relationship. One of the few exceptions is capital per worker, which decreases with rising $Ofs^{EU15}_{jt}$, although revenues seem to increase. In contrast, the predicted values of $Ofs^{EU15}_{jt}$ in column 5 show a more pronounced development. This finding may imply opposing causalities that influence the coefficients, for instance, if rising average wages lead to rising intensities of $Ofs^{EU15}_{jt}$, while rising intensities of $Ofs^{EU15}_{jt}$ reduce the average wage within firms. The predicted $Ofs^{EU15}_{jt}$ exhibits positive impacts on revenues and employment, whereas the capital per worker tends to fall and average wages decline.³³ In general, these correlations suggest that rising $Ofs^{EU15}_{jt}$ comes along with more labor-intensive production.

Columns 4 and 6 reveal that rising intensities of offshoring to the CEECs (hereafter $Ofs^{CEECs}_{jt}$) are associated with growing revenues, increasing export shares, higher average wages, and more capital per worker. However, $Ofs^{CEECs}_{jt}$ is also negatively correlated with the number of employees, while the plants' average wage bill does not show any significant correlation. Production thus appears to become more capital intensive. Combined with higher revenues, this finding suggests that $Ofs^{CEECs}_{jt}$ occurs with boosts in the productivity of businesses, reductions in unit labor costs, and enhanced competitiveness, affirming the results of Dustmann et al. (2014).

Panel B explores the offshoring exposure of jobs with different degrees of complexity and how the share of these job groups correlates with offshoring. It estimates a linear probability model that regresses a binary variable, which indicates workers of the respective tercile of the task distribution, on the regional offshoring measures. Again, columns 1 and 2 show the estimates from a cross-sectional regression of workers in 1995 on future values of offshoring. The coefficients suggest that $Ofs^{EU15}_{jt}$ takes place particularly in industries that intensively use medium-complexity and/or complex labor. Combined with the decreasing share of complex labor within plants in response to increases in $Ofs^{EU15}_{jt}$ (columns 3 and 5), the correlations suggest a substitutability of imported inputs from the EU15 and complex task bundles, as well as a complementarity with simple labor. In contrast, future values of offshoring to the CEECs show no pronounced relation with the frequency of jobs in the various task terciles.³⁴ Over time, the expansion of imported inputs correlates positively with higher relative frequencies of complex jobs, suggesting either complementarity with complex labor or substitutability with simple labor.

4.2 The wage impact of industry exposure to offshoring

Table 4 Industry-Level Regression Results for the Truncated Sample

In an initial assessment, this section seeks to replicate related outcomes by Baumgarten et al. (2013) for comparison purposes. The starting point is Eq. (10), which employs the variation in industry exposure to imported inputs that complement or substitute for various jobs. Table 4 displays the estimated wage elasticities for the truncated 85 percent sample. Column 1 includes the aggregated measure of offshoring $OS_{jt}$ and a full set of worker and endogenous plant controls. The latter controls for industry-specific time trends that go beyond the industry fixed effects. Moreover, the specification includes match fixed effects and year fixed effects to control for time-invariant and unobserved heterogeneity in worker-plant matches and the time trend.³⁵ The offshoring term without interaction shows the wage effect for a virtual job that does not contain any complex tasks, while the associated interaction term indicates the wage changes along the complexity index. The coefficients confirm that performing more complex tasks shields the worker from adverse wage effects of offshoring. Their magnitudes, however, exceed the analogue in, for example, Baumgarten et al. (2013), which may be due to the higher homogeneity in their subsamples of high- and low-skilled workers or by employing match—instead of individual—fixed effects.

Column 2 omits endogenous plant controls and considers increased productivity due to offshoring as an additional channel. This raises revenues and capital per worker, which in turn increases wages. Compared to column 1, the coefficients become more pronounced and suggest an uneven distribution of the gains with respect to job complexity.

The output in column 3 distinguishes the origin of inputs as in Eq. (11). Notably, the coefficients of $Ofs^{CEECs}_{jt}$ become larger in magnitude and statistical significance compared to the aggregated $OS_{jt}$ in columns 1 and 2. It also becomes obvious that $Ofs^{EU15}_{jt}$ features counteracting effects that are not visible in the estimates of $OS_{jt}$, demonstrating the heterogeneous wage effects described by theory and highlighting the importance of distinguishing not only the types of labor but also the types of inputs.

In columns 4–6, I run a two-stage least squares regression to remedy concerns about endogeneity. The instruments are the export supply of intermediate inputs from German offshoring destinations to high-income countries other than Germany $ESI^{HI}_{jtr}$ (7) and ad valorem trade costs of shipping containers from China to Europe.³⁶ Because there is no statistic available to test the instrument strength in the presence of multiple endogenous regressors and heteroskedasticity (Andrews et al., 2019), I rely on a homoskedastic analogue, which are the Sanderson-Windmeijer (SW) F-statistics (Sanderson & Windmeijer, 2016).³⁷ These statistics indicate instrument strength in all 2SLS specifications in Table 4, while a Hanson test does not rule out instrument validity. The instruments are thus assumed to extract the exogenous variation in offshoring and its effect on wages. Compared to column 3, the coefficients become more pronounced, which is the assumed direction and suggests that reverse causality biases the coefficients opposing to the effects from offshoring. Therefore, relatively high wages of complex (simple) jobs lead to higher $Ofs^{EU15}_{jt}$($Ofs^{CEECs}_{jt}$), whereas $Ofs^{EU15}_{jt}$($Ofs^{CEECs}_{jt}$) leads to decreasing relative wages of complex (simple) jobs. Column 5 then omits the plant controls, which yields estimates of offshoring that include the channels of increased productivity (e.g., through increasing revenues). This raises wage differences along the complexity distribution for $Ofs^{EU15}_{jt}$, while $Ofs^{CEECs}_{jt}$ shifts upward with fewer wage differences between jobs of different complexity levels.

The specification in column 6 includes industry-year fixed effects, which absorb any shock at the industry level that is correlated with offshoring and thus also render other industry variables, such as $Ofs^{EU15}_{jt}$ or $Ofs^{CEECs}_{jt}$, perfectly collinear. The interaction term, however, can still be identified as long as there exists within-industry-year variation in task composition. The estimates suggest even more diverse and highly significant wage effects of instrumented $Ofs^{EU15}_{jt}$ along the task complexity index, while the influence of $Ofs^{CEECs}_{jt}$ declines for high levels of job complexity.

The bottom line, however, remains that inputs from CEECs feature a much higher wage effect than inputs from the EU15. This disparity may be due to wage differences between Germany and the EU15 or CEECs and associated firm savings. Such productivity boosts would then foster the positive effect for complementary tasks and the negative effect for substitutable tasks and, hence, address job complexity. The larger the bundle of tasks is, the more protected the worker's total labor input against substitution by imported tasks in the form of intermediate goods. Hence, performing a larger variety of tasks makes it more likely that the worker will be able to compensate for substitution by specializing in other tasks. Regarding labor demand, this mechanism implies, on the one hand, lower (higher) onshore wage elasticities of jobs that are substituted by complex (simple) task-intensive imported inputs. On the other hand, demand shifts toward complementary tasks (to imports) may also result in shifts in the relative job frequency of complex jobs.

Recall that the industry level is important to observe the complementarity and substitutability of offshoring and domestic labor, but it is not necessarily the relevant wage-determining market. Instead, the occupation level seems to be a more suitable unit, since the estimated standard deviation of wages between (within) occupations is 0.1337 (0.2075) and 0.0790 (0.2344) between (within) industries.

4.3 The wage impact of occupational exposure to offshoring

Table 5 First-Stage results of fixed effects instrumental variable regressions

Table 6 Regression results for the truncated wage distribution

The analysis now turns to the baseline regression that analyzes the wage effects of occupation-specific exposures to offshoring. Columns 1–4 in Table 6 are associated with Eq. (5) and estimate wage changes within occupations and worker-plant matches, and relative to the annual plant averages. They do not include other channels emerging from labor demand changes, such as the impact of workers who switch occupations, employers/plants, or unemployment.

The specifications in columns 5–7 replace plant-year fixed effects with controls for plant changes over time and year fixed effects that capture changes driven by the business cycle. These adjustments also indicate whether plants' exports, or endowment of capital per worker, have heterogeneous wage effects with respect to job complexity. If this were the case, the plant-year fixed effect in the baseline regression would not suffice to control for plant heterogeneity other than the effects from offshoring. While most plant-level controls yield wage impacts in accordance with economic theory, the coefficient of the export share in revenues is unexpected and reveals a negative influence. In the presence of spell fixed effects, the negative impact is supposedly associated with domestic demand shocks that affect wages and revenue at the same time. I cluster standard errors at the treatment level, as suggested by Abadie et al. (2017). This means that occupation-year levels account for the heterogeneity in the treatment effects.

In the OLS specifications, up to four endogenous variables remain in the equation: the two regional offshoring terms and their respective interactions with the task index. All 2SLS regressions instrument for them, including additional instruments for overidentification: the region-specific export supply of inputs to other high-wage countries, ad valorem trade costs of shipping containers from China, and their interactions with complexity. The first-stage results (2 × 4) in column 2 (and 6) are shown in Table 5 in the even- (odd-) numbered columns. While the baseline specification does not reject the validity of the overidentifying instruments, the specification with endogenous plant controls rejects a Hansen test, i.e., the orthogonality of the error term to regressors in the second-stage regression. This outcome is not surprising since the test is rejected not only if the overidentifying instruments are invalid but also if the model is incorrectly specified, as is the case with endogenous controls.

Examining the first stage in more detail, the coefficients demonstrate that all instruments exhibit a plausible influence on offshoring and that their impact is significant. Some of the coefficients, however, are less trivial to interpret. For example, the correlation of the export supply of the CEECs with exposure to offshoring to the EU15 is negative, which may imply that the latter is replaced because (suppliers from) the EU15 also offshore production to the CEECs. Its positive and significant interaction term shows that the relationship is less pronounced for complex jobs. Note that the reverse effect does not occur (columns 5 and 6), but the export supply of inputs from EU15 is positively correlated with offshoring to the CEECs. This effect, as described in Baumgarten et al. (2013), and Hummels et al. (2014), is the expected and could be related to trade costs that go beyond the cost of containers. The interacted container costs from China are positively correlated with offshoring to the EU15 and negatively correlated with offshoring to the CEECs. This combination may occur because Germany replaces complex task imports from the EU15 with imports from overseas if trade costs are low or because complex task imports from the EU15 react less sensitively to changes in container costs than other imports. Furthermore, I heuristically test for weak instruments following the procedure developed by Sanderson & Windmeijer (2016).³⁸ The results indicate that all instruments sufficiently explain the respective offshoring terms (instrument strength).

Predicting offshoring in the first stage thus incorporates exogenous variation in offshoring, which facilitates the identification of its causal effect on wages in the second stage. As columns 2 and 6 in Table 6 reveal (compared to columns 1 and 5), the elasticities of offshoring to the EU15 become more pronounced if endogeneity is removed. This change is as expected. It may be due to unobserved shocks that are positively (negatively) correlated with offshoring and have a negative (positive) effect on real wages (omitted variable bias) or—as assumed—reverse causality could cause bias the OLS estimates, e.g., if high wages of complex occupations induce more offshoring activities in the EU15. In contrast, the coefficients of offshoring to the CEECs change only slightly, suggesting a smaller bias due to endogeneity.³⁹

Turning to the interpretation, the baseline regression in column 2 identifies the effects of offshoring to Eastern and Western Europe on wages within occupations, worker-plant matches, and plant-year observations. Specifically, the opposing signs for the two types of offshoring reveal, on the one hand, that both types incorporate substitution and productivity responses on wages and, on the other hand, that these responses show contrasting signs with respect to a job's complexity. While offshoring to the CEECs suggests positive effects on the relative wages of complex jobs, offshoring to the EU15 negatively affects the relative wages of complex jobs.⁴⁰ Although the latter is lower in magnitude, it is an important factor mitigating the relatively strong effects of offshoring to CEECs. Moreover, it is able to reconcile two seemingly contradictory phenomena in the literature: the high substitutability of complex jobs with foreign labor (offshorability) and the positive wage responses of those jobs observed in response to offshoring. While input trade among the EU15 accounts for the bulk of all offshoring activities and moderately lowers wages for complex jobs, the vast expansion of offshoring to CEECs dominates those effects and results in an overall wage divergence between jobs of different complexities.

To eliminate concerns that the opposing effects of region-specific offshoring are caused by multicollinearity, I separately run regressions for each type of offshoring (columns 3 and 4). These specifications, however, are likely to bias the coefficients due to the omitted variables. While the wage elasticity of offshoring to CEECs does not change, the elasticity of offshoring to EU15 countries becomes even more pronounced (compared to column 1). Multicollinearity is therefore unlikely to drive the coefficient signs.

Note that thus far, the specifications eliminate any channel of induced productivity on labor demand either by including plant-year fixed effects or plant controls (e.g., capturing wage increases due to higher revenues). By omitting these, wage elasticities include the productivity effect, which augments the wage impact of imported inputs from the EU15 (column 7), while the productivity effect of imported inputs from the CEECs does not seem to play an important role (similar to column 2). A possible explanation is that a plausible threat to offshore jobs to CEECs changes the bargaining position of workers more severely, which results in a lower labor share in overall income. Note also that in this setup, the coefficient of the interaction term does not exceed the EU15 offshoring term, implying positive average wage effects for complex jobs.

Figure 3 illustrates the baseline results from columns 2 (left, exclusive of the productivity effect) and 7 (right, including the productivity effect) by using the actual evolution of offshoring and by indexing real wages of simple (task = 0.2), medium-complexity (task = 0.5), and complex (task = 0.9) jobs to their values in 1996. Starting with the left graph, it depicts that offshoring to the CEECs increases, ceteris paribus, the average wages of complex jobs by 5.2 percent, while it reduces the average wages of simple jobs by 5.1 percent. If the effect of offshoring to the EU15 is now added, the overall impact of offshoring changes to +4.2 and −3.9 percent, respectively. Adding the productivity effect, the right graph indicates that the wage effects induced by offshoring to CEECs change only slightly; they still negatively (positively) affect workers with simple (complex) task profiles. If both types of offshoring are considered instead, the wage response shifts upwards for all types of workers. Note, however, that relatively simple jobs still suffer slight wage losses, while the discrepancy with the evolution of wages of complex jobs diminishes.

Fig. 3

Wage responses to offshoring shocks by different task complexity sets. Notes: Fig. 3 depicts, ceteris paribus, the evolution of average wages of simple, intermediate, and complex jobs in response to offshoring 1) to the CEECs (thin line) or 2) to the CEECs and EU15 (thick line) from 1996 to 2007. The left panel draws on the estimates in column 2 in Table 6. It shows how offshoring to the EU15 mitigates the amplification of the income gap. The right panel refers to column 7 in Table 6. The evolution of wages now includes the channel of induced productivity from offshoring, where adding offshoring to the EU15 yields more positive wage effects on all types of jobs and a reduced income gap

Thus far, wage regressions have incorporated information up to the 85th percentile of the wage distribution. They ignore truncation, which could affect the estimates for complex jobs, for example, if only a less productive subgroup of the respective occupation is observed. Then, their wages may grow more slowly or decrease more quickly than the actual group average. To obtain information on high wage earners, it is necessary to infer the effect from observable units. I do this in several ways.

As initial evidence, Table 3 (panel B) already indicates differences in labor market outcomes with respect to the type of offshoring and without any truncation. In summary, the correlations suggest that relative labor demand for complex jobs declines when inputs are imported from other high-wage countries and rises when inputs are imported from low-income countries.

In a second exercise, I impute censored entries following the procedure developed by Card et al. (2013) (see B.1.3) and rerun the main specifications from Table 6 on the full wage distribution. Table 7 presents the resulting OLS estimates, which feature the same signs but higher wage elasticities across the complexity distribution. This change is likely due to having a wider range of wages in the sample, which increases the deviations from the mean wage and the covariance with offshoring.⁴¹ The previous tables therefore seem to present rather conservative estimates.

In another approach, I reduce the sample to workers younger than 35 years of age. Their wages are lower for reasons such as having less work experience and tenure, whereas they are not an occupational subgroup that features few productivity-enhancing individuals.⁴² Selecting this subsample leaves 94 percent of the annual wage distribution non-censored. In comparison to the baseline regression, these specifications reveal more pronounced effects on relative wages, suggesting attenuation of the elasticities from the baseline regression.

Table 7 Full Sample with Imputed Wages and the Subsample of Young Workers

5 Further robustness checks

The following section explores alternative specifications and assesses the robustness of the wage effects of offshoring.

5.1 Nonmonotone wage effects along the job complexity measure

The previous results assume a monotone relation between the offshoring terms and the task index and identify winners and losers for each type of offshoring. If offshoring positively affects the demand for some jobs and negatively affects that for others, the estimation assumes that the transition occurs at a given point. From the neighborhood around this point, the wage elasticity further increases towards the poles of the complexity distribution. Such behavior, however, could miss some information since the coefficient of the interaction term could also be driven by wage effects on either less- or more-complex jobs.

It is straightforward to put this possibility to the test by assigning each worker to one of five groups that constitute the quintiles of the complexity distribution. Online Appendix C.1 reports the results of this specification. It does not reject the assumption that the types of offshoring affect wages monotonically with respect to job complexity. Offshoring to the EU15 affects the wages of jobs with either few or many complex tasks. For offshoring to the CEECs, the expanded pattern of wage responses is clearer, revealing a substantial negative and significant impact on rather simple jobs and gradual increases for rather complex jobs.

5.2 The influence of labor market reforms in Germany

A major political debate in Germany persists regarding the economic impact of comprehensive labor market reforms that were introduced between 2003 and 2005, called the Hartz reforms. These reforms will bias the IV estimates if their impact is correlated with the occupational export supply of intermediate goods to other high-income countries and wage changes in Germany. Since the Hartz reforms were mainly intended to lower unemployment and the reservation wages of low-paid occupations, they may have had an adverse influence on the bargaining positions of simple jobs and thereby disturbed the causal identification of offshoring in the above approach. To control for this development, I divide the sample into two periods.

Online Appendix C.2 shows the results for the two successive periods. The estimates suggest that the baseline results are not driven by confounding wage effects of the Hartz reforms.

5.3 Endogeneity of the job complexity index

Another potential threat to identification is the endogenous change of task profiles within occupations. Besides wages and employment, a potential channel through which offshoring impacts the labor market could be a job's adjustment in its typical task bundle. This adjustment could add new and potentially more complex (nonroutine) tasks unevenly across occupations. For example, a given assembly job could be increasingly involved in testing newly developed products or a given engineer job could be less involved in controlling production lines but rather in research and development or management. Any selective changes in the occupations' task profile may alter the rankings underlying the job complexity index and this in turn my cause a bias of the estimated coefficient on the wage effects. If one expects a selectivity that particularly those jobs become more complex that are exposed to high offshoring intensities, then the resulting ranking of occupations in the job complexity index would attenuate the magnitude of the estimated wage effect (e.g., some of the complex jobs may have had a more simple task profile before the sample period and vice versa). In a nutshell, the static complexity index used in the former analysis could be biased from the endogenous ranking of job complexity within the adjustment period. To provide insight on the magnitude of this bias, I thus run the same regressions from 2000 onward. I leave out the second wave of the BIBB-IAB work survey, so job complexity is now measured prior to the exposure to offshoring. As Online Appendix C.3 shows, the results from the baseline regression are robust and a rather conservative estimate of the true effects.

5.4 Alternative instrument

In Tables 4, 6 and 7, I report IV-estimates that use the ESI from Germany's offshoring destinations to other high-income countries. This choice of instruments is a trade-off between instrument strength, so the correlation between ESI and offshoring, and instrument validity, which is the conditional independence of the export supply to wage setting in Germany beyond the channel of offshoring. Thus far, I have decided in favor of the former, since the value chains within Europe feature high correlations. Choosing all high-income countries except for Germany, however, may lacks instrument validity, because the labor markets within Europe could also be interrelated beyond the channel of offshoring. Although it is likely that the multidimensional fixed effects capture most of the other potential channels, particularly time variant effects such as political reforms may still bias the IV-estimates. For this reason, in I now put even stronger emphasis on instrument validity and omit European countries from the measure of export supply of intermediate goods. The remaining high-income export destinations of the ESI are, hence, Australia, Canada, Japan and the USA ($ESI^{Alt}$). I provide the technical details and results of this robustness test in Online Appendix C.4. They exhibit that the results are robust to using the alternative instrument.

5.5 Alternative measures for offshoring

In the previous regressions, offshoring was defined by Eq. (1) as the fraction of imported inputs (from the same industry as the using industry) in the total output of the industry (similar to Baumgarten, Geishecker & Görg, 2013). This measure is designed to capture the importance of imported inputs and it is well suited to compare the international fragmentation of production across industries. Especially, it is able to account for an additional fragmentation of the value chain. However, it also comes with a drawback if applied to time series data, because it mechanically diminishes the effect of enhanced productivity from offshoring. Since this channel is expected to have positive effects on exposed workers' wages, the estimates in the previous regression could be downward biased. To test the magnitude of this bias, I apply an alternative measure for offshoring that considers the share of imported inputs in all inputs from the same industry as the using industry. Online Appendix C.5 documents the technical details and table of results. Overall, the patterns and relative magnitudes of the estimates remain similar to the baseline regressions.

5.6 Alternative measures for task profiles

A final robustness check analyses whether the regression should rely on the tradability of tasks (offshorability) or any other particular characteristic. The selection ranges from a fairly similar index to the measure of routineness and, finally, to the measure of offshorability in Blinder & Krueger (2013). The technical details and results are reported in Online Appendix C.6. In summary, the results seem to be fairly robust to measures that closely measure occupational complexity.

6 Conclusion

The paper distinguishes types of labor by measuring the complexity of jobs. On the production side, it approximates the complexity of imports by considering offshoring to either high- or low-wage destinations. The empirical strategy then identifies wage effects with respect to job complexity and with respect to the type of imported inputs. Due to continuous reductions in European trade costs, the analysis of intra-European value chains is well suited to this subject. Using the most comprehensive dataset for workers in Germany allows the application of multidimensional fixed effects. This approach controls for much of the unobserved heterogeneity. The IV approach solves the problem of the endogenous determination of wages and offshoring by applying time-varying, region-specific instruments. With these tools at hand, the paper reveals wage changes within occupations and worker-plant matches that reach beyond plant-specific shocks.

The key insights of the paper are as follows. First, offshoring to high-income countries, such as the EU15, accounts for the bulk of Germany's imports in intermediate goods and rose substantially after 1996. In absolute terms, this increase is comparable to the increase in offshoring to the CEECs. Second, the characteristics of offshoring destinations have substantially different implications for domestic production. Precisely, the analysis suggests that increasing offshoring to the EU15 entails more labor-intensive production, while increasing offshoring to the CEECs is accompanied by more capital-intensive production. Third, the analysis identifies the causal wage effects of offshoring to high- or low-income countries with respect to job complexity. Complex jobs moderately suffer from wage losses in response to offshoring to the EU15, whereas simple jobs experience wage gains. The paper also finds the opposite impacts for offshoring to the CEECs, and these impacts are of a much higher magnitude. Explicitly, the estimates suggest that offshoring to the CEECs increased the average wage of jobs with high complexity measures of 0.9 by 5.2 percent, while it decreased the average wage of jobs with low complexity measures of 0.2 by 5.1 percent between 1996 and 2007. If one also considers the growth of offshoring to the EU15, the corresponding wage effects are +4.2 and -3.9 percent, respectively.

Finally, the results can reconcile two seemingly contradictory phenomena in the literature: the high substitutability of complex jobs with foreign labor (offshorability) and the positive wage responses of those jobs to offshoring. While input trade among the EU15 accounts for the bulk of all offshoring activities and moderately lowers wages for complex jobs, the vast expansion of offshoring to CEECs dominates those effects and results in an overall wage divergence between jobs of different complexities. These counteracting effects of offshoring not only explain the low and often statistically nonsignificant labor market effects reported in the previous literature but also contribute to the recent debate on the effects of free trade agreements among high-income countries.