Measured Skill Premia and Input Trade Liberalization: Evidence from Chinese Firms

Abstract

Using Chinese firm-level production data, this paper developed a Mincer (Schooling, experience and earnings. Columbia University Press for the National Bureau of Economic Research, 1974)-type approach to investigate the impact of input trade liberalization on firms’ wage inequality between skilled and unskilled workers (or skill premium). When controlling for product-market tariffs in a firm’s industry, we find robust evidence that reduced input tariffs in a firm’s industry are associated with a higher skill premium at firms with more skilled workforces. This effect is more pronounced at ordinary (non-processing) firms. We also provide evidence that reduced input tariffs in a firm’s industry are associated with higher value added and profits at firms with more skilled workforces.

We are deeply grateful to the two anonymous referees and the Editor, Robert Staiger, for their extremely constructive comments and instructions. We thank Andrew Bernard, Binkai Chen, Meredith Crowley, Robert Feenstra, Zhiyuan Li, Kunwan Li, Xiaoning Long, Larry Qiu, Puyang Sun, Wei Tian, Yang Yao, Chuanchuan Zhang, Xiaobo Zhang, Lex Zhao, and various seminar participants for their helpful comments and suggestions. Bo Chen also thanks the National Social Science Foundation of China (Grant 17BJL094) for financial support. Miaojie Yu thanks the National Natural Science Foundation of China (Grants 71573006, 71625007), the National Social Science Foundation of China (Grant 16AZD003), and the Ministry of Education of China (Grant 15JJD780001) for financial support. Zhihao Yu acknowledges the financial support of FPA Research Productivity Bursary from Carleton University.

1 Introduction

Tariffs have declined dramatically worldwide as a result of many rounds of trade negotiations under the General Agreement on Tariffs and Trade (GATT)/World Trade Organization (WTO) (Bagwell & Staiger, 1999). The labor markets in each country have been impacted by the trade liberalization in final-goods and intermediate input sectors. The question of how trade liberalization affects wage inequality between skilled and unskilled workers, especially for developing countries, has once again become one of the research focuses in the international trade literature.

Most of the early studies in the literature follow the HeckscherOhlin model to test whether trade liberalization benefits the abundant factors. According to the Stopler-Samuelson theorem, trade liberalization would mitigate wage inequality between skilled and unskilled labor in the developing countries. However, this theoretical assertion has received little empirical support because most studies find increased skill premium in both developed and developing countries.¹ For example, Feenstra and Hanson (1996, 1999) find that, in the presence of vertical integration and international outsourcing, freer trade could actually increase skill premium in both developed and developing countries^.2

Recent studies in the literature use firm-level data to investigate the impact of globalization on wage inequality but they mainly focus on the impact of export-side trade liberalization (e.g. Bustos, 2011; Frías et al., 2012; Helpman et al., 2017; Verhoogen, 2008). For example, using Brazilian data, Helpman et al. (2017) find that much of the overall wage inequality occurs within sector-occupations, which is mainly driven by wage dispersion between, rather than within, firms. However, the impact of input trade liberalization on firm-level wage inequality is equally important and may also have distinct differences in how the employers might share the surplus with various input factors because of their different bargaining power. In particular, imported intermediate inputs have been found to be crucial for boosting firm productivity in both developed and developing countries such as the United States (Hanson et al., 2005), Indonesia (Amiti & Konings, 2007), India (Goldberg et al., 2010; Topalova & Khandelwal, 2011), and China (Yu, 2015).

The purpose of the present paper is to investigate the impact of input trade liberalization on wage inequality in China and intend to make the following two contributions to the literature. First, investigations on the impact of input trade liberalization on wage inequality in developing countries usually rely on industry-level wage data, household survey data, and the Gini coefficient as a proxy for income inequality (e.g., Beyer et al., 1999).³ For example, using urban industrial survey data, Han et al. (2012) found that widening wage inequality in China was strongly associated with China’s accession to the WTO in 2001.⁴ In the present paper, we use Chinese firm-level production and customs’ trade data to investigate the impact of tariff reductions for imported inputs on firm-level skill premium in China. To our knowledge, this is the first paper to investigate how import trade liberalization affects firm-level skill premium for manufacturing firms in China, the largest developing country in the world. The study could enrich our understanding of the sources of China’s growing income inequality from the wage differentials at the firm levels.⁵

Second, a major challenge to investigate firm-level wage inequality between skilled and unskilled labor in China, as in most developing countries, is lack of data for direct firm-level wages for skilled and unskilled workers. To overcome this major obstacle, the current paper has developed a method of constructing firm-level skill premium from a firm’s average wage and share of skilled labor. Together with a Mincer (1974)-type regression, we are able to estimate the impact of input trade liberalization on firm-level skill premium in China. This method can be applied to other research facing similar data limitations.

Using firm-level production and transaction-level trade data from China, we find that, when controlling for product-market tariffs in a firm’s industry, the reduced input tariffs in a firm’s industry are associated with a higher skill premium at firms with more skilled workforces. This effect is more pronounced at ordinary (nonprocessing) firms. Compared with processing importers, ordinary importers respond more forcefully to input trade liberalization in their wage schedule. Our main finding that input trade liberalization increases firm-level skill premium in China is robust for all three regions (east, central and west) in China, as well as for different measures of wage inequality, different empirical specifications and data spans. By contrast, output trade liberalization, which is measured by product-market tariffs reduction, show opposite signs of the effect of import tariffs on the skill premium at firms with more skill-intensive production. However, the product-market tariff effects are not robustly significant across specifications.

Inspired by the literature on “fair wages” (e.g. Egger & Kreickemeier, 2012), we also provide an interpretation for our main finding that input trade liberalization leads to an increase in firm skill premium. If skilled workers have greater bargaining power with their employers than unskilled workers, incomes of the skilled workers shall be more closely linked to firms’ economic profits but the incomes of unskilled workers shall be more in line with those of other firms in the same industry. Thus, a fall in input tariffs increases the firm’s value-added, which in turn raises the firm’s skill premium because the skilled labor commands a larger proportion of the incremental surplus than the unskilled labor. We also provide some evidence for our conjecture.

In addition to the literature discussed earlier, our paper is also closely related to the studies on how import trade liberalization affects skill compositions and factor returns. For example, using data from multinational companies, Biscourp and Kramarz (2007) examine the impact of offshoring on plant-level skill composition in France. Similarly, Becker et al. (2013) investigate the impact of offshoring on firm-level task composition and wages in Germany. Amiti and Davis (2011) is another influential study that investigates the impact of output and input tariff reductions on wages. In particular, they find that a reduction in-input tariffs raises wages at import-using firms relative to those using only domestic intermediate inputs. However, these studies do not focus on firm wage inequality, or skill premium.

The rest of the paper is organized as follows. Section 2 describes the data and introduces the econometric methods to measure firm skill premium and the empirical specifications of Mincer regression. Section 3 presents the main empirical evidence, offers robustness checks and discussion on the possible mechanism. Section 4 concludes.

2 Data, Measures, and Empirics

2.1 Data

To investigate the impact of input trade liberalization on firms’ skill premium, our analysis uses the following three disaggregated panel data sets: firm-level production data compiled by China’s National Bureau of Statistics (NBS), production-level trade data maintained by China’s General Administration of Customs, and China’s import tariff (ad valorem) data at the HS 6-digit level, maintained by the World Integrated Trade Solution (WITS) database of the World Bank.

China’s NBS conducts an annual survey of industrial firms (ASIF) with two types of manufacturing firms: all state-owned enterprises (SOEs) and non-SOEs whose annual sales exceed RMB 5 million (or equivalently $725,000). The sample used in this paper has ~ 230,000 manufacturing firms per year, varying from 162,885 firms in 2000 to 301,961 firms in 2006. On average, the sample accounts for more than 95% of China’s total annual output in the manufacturing sectors.⁶ The data set covers more than 100 accounting variables and contains all of the information from the main accounting sheets, which includes balance sheets, loss and profit sheets and cash flow statements.

Given its rich information, the firm-level production data set is widely used in research by, among others, Cai and Liu (2009), Brandt et al. (2012), and Feenstra et al. (2014). However, the data set has two limitations for our research purpose. The first one is common: some unqualified firms are wrongly included in the data set, largely because of mis-reporting or false recording. Thus, following Feenstra et al. (2014), we keep the observations in our analysis according to the requirements of the Generally Accepted Accounting Principles (GAAP).⁷ Accordingly, the total number of firms covered in the data set was reduced from 615,951 to 438,165, and approximately one-third of the firms were removed from the sample after the rigorous filter was applied. The drop in the percentage of sales is only around 25%. Thus, the drop in sales is smaller, since larger firms meet the GAAP more frequently. This suggests that larger firms that follow GAAP may generate larger rents per worker, so they have a large surplus to share and the measured effect of reduced input tariffs might be an upper bound.

The second limitation of the data is specific to the present paper. The data set does not separate wages for skilled and unskilled labor. Furthermore, the numbers (i.e. the share) of skilled and unskilled workers are only available for 2004. To overcome this problem, we conduct our baseline test on cross-firm data for 2004. Then we carry out panel data tests that include other years by multiplying the skilled labor shares in 2004 by the change in the skilled labor share (relative to 2004) at the provincial level. To ensure the precision of our estimates, we exclude the pure trade intermediaries (that do not have production activities) from the sample in all the estimates. The trade intermediaries are identified according to the same procedures as in Ahn et al. (2011).

Finally, we use customs’ data to match with the firm-level production data so that we are able to identify each firm’s importing and processing status. As introduced in Feenstra et al. (2014), the production-level trade data maintained by China’s General Administration of Customs include a large variety of information such as each trading firm’s importing (or exporting) status and processing (or non-processing) status. Such information is essential for us to conduct our empirical estimations, which will be discussed shortly.

2.2 Measures

This subsection starts by introducing the index of input trade liberalization, and then focuses on constructing firm-level measured skill premium because the data sets do not directly provide firm-level wages for skilled and unskilled labor.

2.2.1 Measures of Input Tariffs

Inspired by Amiti and Konings (2007) and Topalova and Khandelwal (2011), we construct the industry-level input tariffs, ${IT}_{j}$ as follows:

$${IT}_{j}=\sum_{n}\left(\frac{{input}_{nj}^{2002}}{\sum_{n}{input}_{nj}^{2002}}\right){\tau }_{n}$$

(1)

where IT_j denotes the industry-level input tariffs facing firms in industry j in 2004 and t_n is the tariff on input n in 2004. The weight in parentheses is the production cost share of input n in industry j.

We use China’s Input-Output Table of 2002 to construct the weight because NBS reports the Input–Output Table every 5 years and our data are for 2004. In the spirit of Bartik (1991), we use the input-output matrix from 2002 to compute the relevant weighted industry input tariffs since the weight in 2002 reflects the initial conditions prior to China’s tariff cuts in 2004.⁸ The industrial input tariffs are obtained as follows. First, since there are 71 manufacturing sectors reported in China’s Input–Output Table (2002) and only 28 manufacturing sectors reported in the Chinese Industrial Classification (CIC), we start by making a concordance between the Input–Output Table and the CIC sectors. Second, we match the CIC sectors with the International Standard Industrial Classification (ISIC, rev. 3).⁹ Third, we make another concordance to link the ISIC and HS 6-digit trade data, where we can find the corresponding tariffs from the WITS database. Fourth, we calculate the industry-level tariffs that are aggregated to the CIC sector level.¹⁰ Since simple-average tariffs cannot take into account the difference of the importance of imports, we consider the following weighted industrial tariffs:

$${\tau }_{n}=\sum_{k\in n}\left(\frac{{m}_{k}}{\sum_{k\in n}{m}_{k}}\right){\tau }_{k}$$

(2)

where m_k is the import values for product k in CIC 2-digit industry n in 2004. Finally, we calculate the industry-level input tariffs using Eq. (1). The industry-level output tariff for industry n in 2004 is also obtained from Eq. (2).

To see how the input tariff reductions affect firms’ skill premium, we examine the evolution of China’s trade liberalization throughout the sample period. Table 1 reports the mean and standard deviation for this key variable by spreading the sample from 2000 to 2006. As shown in Table 1, the average industry input tariffs were cut in half, from 15.73% in 2000 to 7.71% in 2006, and their standard deviation also dropped by about a half over the same period. The industry input tariffs were around half of their initial levels in 2000, before the WTO accession. Finally, the industry input tariffs in 2004 were also lower than the corresponding industry output tariffs, as shown in Table 2.

Table 1 China’s industrial input tariffs

Table 2 Summary statistics of key variables (2000–2006)

2.2.2 Measures of Skill Premium

The skill premium is defined as ${s}_{it}=({w}_{it}^{s}-{w}_{it}^{u})/{w}_{it}^{u}$ for the skilled wages $({w}_{it}^{s})$ and unskilled wages $({w}_{it}^{u})$.¹¹ Given the share of firm i′ s skilled workers $\left( {\theta_{it} } \right)$, the firm average wage $\left( {\overline{{w_{it} }} } \right)$ can be written as $\overline{{w }_{it}}={\theta }_{it}{w}_{it}^{s}+(1-{\theta }_{it}){w}_{it}^{u}$ or, relative to the unskilled wages, $\overline{{w }_{it}}$/${w}_{it}^{u}$ = 1 + ${\theta }_{it}{s}_{it}$. Hence, the log term of the average wage is:

$$ln \overline{{w }_{it}}= ln{w}_{it}^{u}+ln(1 + {\theta }_{it}{s}_{it})$$

(3)

When ${\theta }_{it}{s}_{it}$ is small, we can omit the higher-order terms and have $(1 + {\theta }_{it}{s}_{it})\approx {\theta }_{it}{s}_{it}$. Therefore,

$$ln \overline{{w }_{it}}\approx ln{w}_{it}^{u}+{\theta }_{it}{s}_{it}$$

(4)

The key advantage of Eq. (4) is that it gives rise to a plausible Mincertype regression for our empirical estimation. The trade-off is that, if ${\theta }_{it}{s}_{it}$ is not small enough, our Mincer-type regressions are not precise enough to interpret the economic magnitudes of the estimated coefficients. Hence, in the rest of the paper, economic interpretation should be focused on the sign, rather than the magnitude, of our estimates.¹²

Table 2 reports the summary statistics for the key variables used in our estimations. In the firm data set, information on firms’ skilled labor share is available only for 2004, although firms’ average wages are available for 2000–2006. Since firms’ skill share is crucial in Specification (4), we use the cross-section data for 2004 to conduct the main analysis and a panel sample for 2000–2006 for robustness checks only. Since the firm-level data set provides employment information on skilled and unskilled labor only for 2004, we use a proxy for the skilled labor share for all other years. To obtain the proxy ($\widehat{{\theta }_{it}}$), we multiply the skilled labor share in 2004 (${\theta }_{i,2004}$) by the provincial skilled labor share (${\eta }_{rt}$) in all years, using 2004 as the base year: $\widehat{{\theta }_{it}}\equiv {\eta }_{rt}{\theta }_{i,2004}$. Table 2 reports the mean and standard deviation of the key variables for the samples for 2004 and 2000–2006.

Three variables in Table 2 relate to wage information. The first is firm average wage, which is reported from the data sets directly. The second is the measured wage premium (${\mu }_{i}$), which is defined as firm I’s log wage relative to that of the average firm in industry j and region r (to be discussed in details in the next section). The last wage variable is the measured unskilled wage. Since the annual survey of industrial firms does not provide firm-level unskilled wages, we define the measured unskilled wage as the minimum level of firm wages in each (3-digit) industry-province pair based on the following two observations. First, as shown in Table 2, firms’ average wages are significantly positively correlated with the skill share,¹³ but the mean of measured unskilled wages is much lower than that of the firms’ average wage (around 15%). Second, according to Anwar and Sun (2012), wages of unskilled workers in China are actually different across industries and provinces, especially after 2004. As a robustness check, however, we also use an alternative measure of the unskilled wages for our estimations. Finally, the firm-level data set for 2004 reports five education levels: graduate (and above), university, college, high school, and below middle school. As in most studies, we define skilled workers as employees with a college degree or higher.

2.3 Mincer Empirical Specification

Let us suppose that firm i^rs skill premium, s_i, takes a linear form

$$ s_{it} = \mathop \sum \limits_{p = 0}^{P} \gamma_{p} \chi_{it}^{p} + \varepsilon_{it} $$

(5)

where x^p denotes a vector of predictors. From Eqs. (4) and (5), we obtain the following. Mincertype empirical specification:

$$ \begin{aligned} \ln \overline{{w_{it} }} & = \gamma_{0} + \gamma_{u} \ln w_{it}^{u} + \gamma_{0} \hat{\theta }_{it} + \gamma_{1} \left( {\hat{\theta }_{it} IT_{jt} } \right) + \gamma_{2} \left( {\hat{\theta }_{it} IT_{jt} } \right)IM_{it} \\ & \;\;\;\; + \gamma_{3} \left( {\hat{\theta }_{it} PT_{jt} } \right) + \gamma_{4} \left( {\hat{\theta }_{it} PT_{jt} } \right)FX_{it} + \gamma_{5} \left( {\hat{\theta }_{it} FX_{jt} } \right) \\ & \;\;\;\; + \gamma_{6} \left( {\hat{\theta }_{it} IM_{jt} } \right) + \gamma_{7} IT_{jt} + \gamma_{8} PT_{jt} + \gamma_{9} IM_{it} \\ & \;\;\;\; + \gamma_{10} FX_{it} + \gamma_{11} \mu_{it} + \gamma_{12} \hat{\theta }_{it} \mu_{it} + \gamma X_{ijt} \\ & \;\;\;\; + \delta_{i} + \delta_{jr} + \delta_{t} + \varepsilon_{it} \\ \end{aligned} $$

(6)

where the error term is defined as $\varepsilon_{it} \equiv \widehat{{\theta_{it} }}\varepsilon_{it}$. The main regressors in this Mincer regression include three sets of variables: (i) we include unskilled wage (ln(${w}_{it}^{u}$)), measured skilled labor share (${\widehat{\theta }}_{it}$) and its interaction with input tariffs (IT_jt) and output tariffs (PT_jt)¹⁴; (ii) we also include import dummy (IM_it) (export dummy, FX_it) and its interaction with tariffs; (iii) We also include the own terms and their interaction terms with skill share of firm-level controls X_ijt such as firm ownership (state-owned enterprise, foreign firm, or private firm), firm size (poxied by firms’ log sales), and firm productivity; finally, (iv) in addition to firm-specific fixed-effects (${\delta }_{i}$), interacted industry-region fixed-effects (${\delta }_{jr}$), and year-specific fixed effects (${\delta }_{t}$), we also include firm wage premium, defined as ${\mu }_{it}\equiv ln \overline{{w }_{it}}-{\sum }_{i\in I(jr)}^{N}(ln \overline{{w }_{it}})/|\mathrm{Jr}|$ where |Jr| is the cardinality of the set of it i ∈ I(jr) it firms in industry-region pair jr, and its interaction with. Firms’ skill share in the regression.

Among the regressors, there are five important points that are worth noting. First and foremost, among the set of predictors, the most important variable of interest is the average intermediate input tariffs in industry j (IT_j) that firm i is associated with. If the coefficient c₁ in Eq. (6) is negative and statistically significant, it suggests that input trade liberalization would increase firm skill premium. It is also reasonable to anticipate that the impact of input trade liberalization on skill premium would be stronger for ordinary (i.e., non-processing) importing firms, since processing imports have already enjoyed the special treatment of free duty (Yu, 2015) and hence would be less impacted by a further input trade liberalization. Thus, we expect that c₂, the triple interaction term among skill share, intermediate input tariffs, and the importer indicator, should be negative. By constrast, another triple interaction term among skill share, intermediate input tariffs, and the processing indicator (not shown in the Eq. (6)) is expected to be positive.

Second, we include the industry average output tariff (PT_j) and its interaction with the firm export indicator as control variables for the reasons as follows. After its accession to the WTO, China cut not only its intermediate input tariffs, but also its final output tariffs (see Yu, 2015, for a detailed discussion). It would be expected that the impact of output trade liberalization on wage inequality may be different between exporting firms and non-exporting firms. Thus, the interactions of output tariffs with firm-level exporting indicators are introduced for that purpose (see Biscourp & Kramarz, 2007; Verhoogen, 2008). Of course, skill premium in exporting (importing) firms may be affected through channels other than trade liberalization. We thus also include firms’ own exporting and importing indicators in the regressions.

Third and equally important, the regression Eq. (6) requires panel data. However, we have recorded data on the share of skilled labor only for year 2004, and a proxy for the share of skilled labor for other years: $\widehat{{\theta }_{it}}\equiv {\eta }_{rt}{\theta }_{i,2004}$, which is the skilled labor share in 2004 (${\theta }_{i,2004}$) multiplied by the provincial skilled labor share (${\eta }_{rt}$) in all other years by using 2004 as the base year. The limitation of using the above panel-data estimation is that the within-firm variation generated by the interaction terms of $\widehat{{\theta }_{it}}$ are mainly from the ${\eta }_{rt}$ portion. Thus, we will first use 2004 data and the following baseline cross-section regression for our estimation:

$$ \begin{aligned} \ln \overline{{w_{i} }} & = \gamma_{c} + \gamma_{u} \ln w_{i}^{u} + \gamma_{0} \theta_{i} + \gamma_{1} \left( {\theta_{i} IT_{j} } \right) \\ & \;\;\;\; + \gamma_{2} \left( {\theta_{i} IT_{j} } \right)IM_{i} + \gamma_{3} \left( {\theta_{i} PT_{j} } \right) \\ & \;\;\;\; + \gamma_{4} \left( {\theta_{i} PT_{j} } \right)FX_{i} + \gamma_{5} \left( {\theta_{i} FX_{i} } \right) \\ & \;\;\;\; + \gamma_{6} \left( {\theta_{i} IM_{i} } \right) + \gamma_{7} IT_{j} + \gamma_{8} PT_{j} \\ & \;\;\;\; + \gamma_{9} IM_{i} + \gamma_{10} FX_{i} + \gamma_{11} \mu_{i} \\ & \;\;\;\; + \gamma_{12} \theta_{i} \mu_{i} + \gamma X_{i} + \delta_{jr} + \varepsilon_{i} . \\ \end{aligned} $$

(7)

Fourth, ${\mu }_{it}$ is firm i’s log wage relative to the average firm in industry j and province r. These we premia (or discounts) can come from different skill composition of firm i’s workforce, or the different surplus that firm i generates. It is important to emphasize that this variable plays an important role here. It helps us properly control for between-firm skill premium (e.g. Amiti & Davis, 2011; Egger & Kreickemeier, 2009; Helpman et al., 2017). In the cross-section regression in Eq. (7), with proper region-industry fixed effects, the second component ${\sum }_{i\in I(jr)}^{N}(ln \overline{{w }_{it}})/|\mathrm{Jr}|$ of the measured between-firm wage premium (${\mu }_{it}$) should be fully absorbed into the industry fixed effects. Thus, the OLS estimator would then exhibit a coefficient of the variable ${\mu }_{it}$ close to unity. The interaction term ${\theta }_{i}{\mu }_{i}$ is also needed for our Mincer-regression specification.

Finally, our empirical specifications implicitly draw on theory suggested by Helpman et al. (2010). By treating multiple skill groups in the firm-level framework, the regression residuals will depend on (i) the tightness of the local labor market in a province-industry pair, (ii) the locally available skilled workers in an industry and location, (iii) firms’ anticipated performance and associated wage offers, and (iv) any firm-specific shocks to the wage bargaining or screening technology (Blaum et al., 2015; Helpman et al., 2017). Thus, we add the following three sets of dummies in the regressions. First, we include province-specific fixed effects to control for province invariant but unobservable factors (such as export subsidy rates, etc.). Second, we include 2-digit industry-specific fixed effects, which control for industry-invariant factors such as industrial capital intensity. Third, we allow for a full set of interacted industry-province dummies to absorb local labor market conditions. The remaining identifying assumption is the idiosyncratic effect ${\varepsilon }_{i}$∼N(0, ${\sigma }^{2}$),which takes into account firms’ anticipated performance and firm-specific shocks that do not differentially affect individual skill groups.

Some studies have investigated whether more productive firms use more skill-biased technology (e.g., Bustos, 2011; Verhoogen, 2008). It is possible that trade liberalization induces the most productive firms to adopt skill-biased technology or upgrade product quality, and hence increases the demand for skilled labor for these firms. If such a multi-collinearity problem is a big concern, our data should exhibit a strong negative correlation between input tariffs and the skill share. However, this is not the case for our sample as the simple correlation in 2004 cross-section data between industrial input tariffs and the skill share is small (− 0.11). Moreover, the simple correlation in the whole sample for 2000–2006 is even smaller in absolute value (− 0.06). The low correlations suggest that the change of firms’ skill share is not sensitive to the change of trade liberalization, at least in our current sample.

3 Estimation Results

3.1 Baseline Mincer Regressions

Table 3 presents the baseline results for the cross-section empirical specification Eq. (7). Since the firm-level data set does not report firms’ import status, Table 3 does not include the importer indicator. Columns (1) and (2) are a single regression in which column (1) reports the own coefficient of each regressor whereas column (2) reports its corresponding coefficient interacted with the skill share. From column (2), the coefficient of industry input tariffs interacted with firm skill share, the key variable of interest, is negative and statistically significant, suggesting that input trade liberalization tends to increase skill premium. Sheng and Yang (2016) provide evidence that foreign firms in China attract more skill-intensive production, which in turn would raise firms’ skill premium. Thus, we include the interactions of skill share with the foreign indicator and with the SOE indicator in the regression.¹⁵ The positive sign of the coefficient of the foreign indicator ascertains the finding in Sheng and Yang (2016). We also include firm size (proxied by firms’ log sales) and firm total factor productivity (measured by the augmented Olley and Pakes (1996) approach, as suggested by Yu, 2015). Exporting firms may have their own channels affecting the skill premium. We thus interact the skill share with the exporting indicator in column (2). Consistent with most of the previous studies, we find that the skill premium is higher for larger firms, more productive firms, and exporting firms.

Table 3 Mincer regression using matched data for 2004

It is also reasonable to anticipate that firms of different sizes may respond differently to input tariffs. Therefore, we run another regression with results jointly shown in columns (3) and (4). Specifically, we include a triple interaction term among the skill share, input tariffs, and firms’ log sales. Importantly, the skill share is an important predictor in itself, beyond size (log sales), and that the skill share prediction is slightly weaker at larger firms. Because the coefficient of the novel triple interaction term among input tariffs, skill share, and log sales has a different (i.e., positive) sign than that of the interaction term between input tariffs and skill share, we take a further step to evaluate their net effect at the sample mean. Overall, the net effect of input tariffs on firm average wages is still negative since (− 0.181 + 0.013 × 9.94) × 0.449 < 0 given that the sample mean of log of firm sales is 9.94 whereas that of firm skilled share is 0.449 as seen from Table 2. Thus, the counteracting effect generated by the new triple term with log sales does not overwhelm our previous main finding.

As recognized by Cai (2010), China’s labor force generally migrates from the inland (i.e., western and middle) provinces to the costal (eastern) provinces. It is reasonable to expect that firms have different wage premiums in the different regions. We thus classify all 31 provinces into three regions: east, middle, and west. In the single regression as reported in columns (3) and (4), we also take a step further to control for region-specific fixed effects and industry-specific fixed effects to take into account local market tightness (as discussed earlier). In addition, we also include a full set of interacted industryregion dummies. With such rich sets of fixed effects controlled, the coefficient of input tariffs interacted with skill share—our main interest in the estimation—still remains negative and statistically significant.

Our main empirical specification Eq. (6) also permits a regional analysis by grouping the sample for 2004 into three regions: east, central, and west. We first split the entire national sample into 31 provinces and then repeat the Mincer regression, similar to columns (3–4) of Table 3, for the east region, central region and west region. Results are reported in columns (1–2), (3–4), and (5–6) of Table 3, respectively. In each regression, we control for the interacted province and industry fixed effects. According to the China Regional Statistical Yearbook, the eastern region includes fifteen provinces, the central includes six provinces, and the western region includes the rest of the provinces.¹⁶ Thus, the regional regression for the eastern region has the largest number of observations, followed by the west region, and then by the central region. In the three regressions shown in Table 3, the interaction terms between industry input tariffs and skill share are all negative and highly statistically significant.¹⁷ Thus, our main finding that reduced input tariffs in a firm’s industry are associated with a higher skill premium at firms with more skilled workforces is robust for all three regions in China.

3.2 Mincer Regressions Using Matched Sample

Table 3 use the firm-level data set for 2004 to conduct the regressions. The advantage of using only this data set is that it contains all manufacturing firms. Yet, the data set does not contain information on firms’ import status. To overcome this data challenge, we match the ASIF data set with the product-level customs data to perform similar Mincer-type regression in Table 3.¹⁸

Columns (1) and (2) in Table 3 are a single regression with industry-region fixed-effects in which column (1) reports the own coefficient of each regressor whereas column (2) reports the corresponding variables interacted with skill share. Different from estimates in Table 3, we include firms’ importing status in estimates of Table 3 since this variable can better capture firm’s exposures to globalization. The regression shown in columns (1) and (2) includes the own variable of firms’ importing indicator and its interaction with the skill share. The coefficient of industry input tariffs interacted with the skill share is negative and statistically significant, suggesting that reduced input tariffs in a firm’s industry are associated with a higher skill premium at firms with more skilled workforces. In addition, the regression includes a triple interaction term among the importer indicator, skill share, and industry input tariffs. The negative, though insignificant, triple interaction term hints that importers might respond more forcefully to input trade liberalization in their wage schedule.

Furthermore, import processing firms may behave differently from ordinary firms, as suggested by Dai et al. (2016). By definition, import processing firms are firms that import raw material or intermediate inputs and then, after local processing or assembly, export the value-added final goods (Feenstra & Hanson, 2005). A processing indicator is defined as one (1) if a firm has any processing imports and zero (0) otherwise. As processing imports have zero import tariffs (Yu, 2015), the effect of input trade liberalization on firms’ skill premium is expected to be less pronounced for an industry with many import processing firms.

We run another regression which is jointly reported in columns (3)–(5). As before, column (3) reports the own coefficients of regressors whereas column (4) shows the coefficients of their interaction with skill share. Column (5) reports the coefficient of triple interaction among input (output) tariffs, skill share, and processing indicator. Similar to our previous findings, reduced input tariffs in a firm’s industry are associated with a higher skill premium at firms with more skilled workforces, because the coefficient between industry input tariffs and skill premium is negative and statistically significant. The novel finding is that the coefficient of the triple interaction term between industry input tariffs, skill share, and processing indicator is positive and statistically significant, suggesting that the effect of input trade liberalization on firms’ skill premium is more pronounced for non-processing (i.e., ordinary) firms.

To see this more precisely, we can take one step further to evaluate the net effect of input trade liberalization on firm wages. As shown in the single regression of columns (3)–(5), there are three types of firms in the regression: non-importing firms, ordinary importers, and processing importers. The net effect of input tariffs on firm average wage for non-importing firms is 0.0290.088 $\times$ 0.45 < 0 given that the sample mean of firm’s skill share is 0.45 and the own coefficient of input tariffs is 0.029 whereas that of the interaction between input tariffs and skill share is − 0.088. Similarly, the net effect for ordinary importers (i.e., importer indicator equals one) is 0.029 − (0.088 + 0.017) × 0.45 < 0 given that the coefficient of the triple interaction term among input tariffs, skill share and import indicator is -0.017. These two results are consistent with our previous main finding that a fall in industry input tariffs is associated with a higher skill premium at firms with more skilled workforce.

By contrast, the net effect of input tariffs on firm wages for processing firms is positive since it equals 0.029 + (0.137 − 0.088 − 0.017) × 0.45 > 0 given that the coefficient of the triple interaction term among input tariffs, skill share and processing indicator is 0.137. The finding that the sample-mean effect for processing exporters is overturned is also intuitive. Processing imports in China enjoy a special treatment of free duty (see, e.g., Yu, 2015). Thus, further import tariff reductions on processing input encourage processing exporters to switch to ordinary exporters over time, as found by Brandt and Morrow (2017), which in turn lower the employment demand for processing exporters. Accordingly, the average wages for processing firms fall, as shown in the regression of Columns (3)–(5) of Table 3.

3.3 Estimates Using Panel Data

So far, we have used data only for 2004 to estimate the Mincer regressions, because data on firms’ skill shares are only available for census year of 2004. The empirical specifications are useful for understanding cross-section firms’ skill premium. To gain a better understanding on the variation of within-firm skill premium in response to input trade liberalization, in this section we make an effort to use the panel data for the period of 2000–2006.

Since data on the share of skilled labor are available only for year 2004, to compute a proxy for the skilled labor share for all other years from 2000 to 2006, we multiply the skilled labor share in 2004 by the provincial skilled labor share in all the other years using 2004 as the base year. In addition, industry input and output tariffs are now calculated using the Input–Output Table for 1997 to obtain the corresponding weights because the information in the Input–Output Table of 1997 reflects the initial conditions prior to China’s trade liberalization in 2001 (Bartik, 1991).

As data on the share of skilled labor are unavailable for years other than 2004, we compute a proxy for the skilled labor share (hit) for all other years from 2000 to 2006 by multiplying the skilled labor share in 2004 with the provincial skilled labor share in all other years using 2004 as the base year. Equally important, industry input and output tariffs are now calculated using the Input–Output Table for 1997 to calculate the corresponding weights, as the weights in 1997 reflect the initial conditions prior to China’s trade liberalization in 2001, as suggested by Bartik (1991).

With cross-section data in 2004, Table 3 has already demonstrated that the results for empirical specification with both own coefficients and coefficients interacted with skill share for each regressor are very close to those without own coefficients. Since the latter specification follows Mincer regressions more closely, in the panel-data analysis we only report those empirical results of estimation with the coefficients interacted with firms’ skill share.

Column (1) of Table 4 reports the Mincer regression results by using the 1997 Input–Output Table and controlling year-specific fixed-effects, industry-specific fixed effects, and region-specific fixed effects, respectively. The estimation results are very close to their counterparts in the last two columns of Table 3. The coefficient of industry input tariffs interacted with firms’ skill share is negative and statistically significant, indicating that input trade liberalization increases firm’s skill premium over time. Similar to the estimation results shown in column (6) of Table 3, the coefficient of output tariffs interacted with the skill share is positive, for the same reason discussed earlier. Estimates in column (2) take a step further to run a more parsimonious regression by controlling the interacted industry and region fixed effects. All regressors have very similar coefficients to their counterparts in column (1).

Table 4 Mincer regression using the 1997 IO table (2000–2006)

Finally, it is possible that firms may take more time to respond to tariff reductions in their wage schedule. In our last enrichment, column (3) of Table 4 instead uses firms’ past (i.e., one-year lag) export status and past performance (sing log sales or total factor productivity as a proxy). The estimation results for all the variables in column (3), with some variables in one-lag period are close to their counterparts in column (2) when all variables are in the current period. In all cases, the coefficients of industry input tariffs are found to be negative and statistically significant for all the regressions.

3.4 Endogeneity Issues

In the previous estimations, input trade liberalization was considered as exogenous. However, tariff formation could be endogenous in the sense that skill premium could have a reverse effect on tariff changes. With widening skill premium, unskilled workers could blame free trade policies and form labor unions to lobby the government for temporary trade protection (Bagwell & Staiger, 1990, 1999; Bown & Crowley, 2013). Although this happens in developed countries like the United States (Goldberg & Maggi, 1999) and in some developing countries like Turkey (Gawande & Bandyopadhyay, 2000), it is less likely to happen in China because labor unions in China are symbolic organizations. In addition, if these types of political factors are time invariant, they should have been accounted and statistically controlled for by the fixed-effect panel estimates in Table 4 (Goldberg & Pavcnik, 2007). However, if they are time variant, the estimations of the related Mincer regressions in Table 4 would be biased.

Moreover, if the residual in Eq. (6), e_it, is related to the firm’s measured skill share (${\widehat{\theta }}_{it}$), the estimated coefficients will be biased. As a robustness check, below we use the instrumental variables (IV) approach to address the potential endogeneity issues. If the negative reverse causality is a main source of endogeneity issue, we should expect that the key estimated coefficient for the interaction term between input tariffs and skill share under the two-stage least square (2SLS) approach should be greater than its counterpart under the OLS approach.

It is challenging to find an ideal instrument for tariffs. Inspired by Trefler (2004) and Amiti and Davis (2011), we use the one-year lag of industry input tariffs as the instrument of the first difference in industrial input tariffs. The economic rationale is that lagged input tariffs are less likely to influence the time difference of input tariffs (Trefler, 2004). In particular, we consider the following first-difference Mincer regression:

$$ \begin{aligned} \Delta \ln \overline{{w_{it} }} & = \gamma_{c} + \gamma_{u} \Delta \ln w_{it}^{u} + \gamma_{0} \Delta \hat{\theta }_{it} + \gamma_{1} \Delta \left( {\hat{\theta }_{it} IT_{jt} } \right) \\ & \;\;\;\; + \gamma_{2} \Delta \left( {\hat{\theta }_{it} IT_{jt} } \right)IM_{it} + \gamma_{3} \Delta \left( {\hat{\theta }_{it} PT_{jt} } \right) \\ & \;\;\;\; + \gamma_{4} \Delta \left( {\hat{\theta }_{it} PT_{jt} } \right)FX_{it} + \gamma_{5} \Delta \left( {\hat{\theta }_{it} FX_{it} } \right) \\ & \;\;\;\; + \gamma_{6} \Delta \left( {\hat{\theta }_{it} IM_{jt} } \right) + \gamma_{7} \Delta IT_{jt} + \gamma_{8} \Delta PT_{jt} \\ & \;\;\;\; + \gamma_{9} \Delta IM_{it} + \gamma_{10} \Delta FX_{it} + \gamma \Delta X_{jt} + \delta_{i} \\ & \;\;\;\; + \delta_{jr} + \delta_{t} + \varepsilon_{it} \\ \end{aligned} $$

(8)

Accordingly, the regressand and all regressors in Table 5 are in the first difference. Columns (1) and (2) are a single OLS regression in which IV reports the coefficients of the own one-lag industry input tariffs and its interaction with firm skill share using the first difference in industry input tariffs and its interaction with firm skill share as the regressands. Once again, the interaction term between skill share and industry input tariffs is negative and statistically significant, which is consistent with our previous findings. Finally, to show that our 2SLS estimation results are robust to the inclusion of the own terms of the regressors, we run another single estimation by abstracting away the own coefficients of related regressors, which is jointly reported in columns (3) and (4). Similarly, the regression in columns (3) and (4) use the one-lag industry input tariffs interacted with firm skill share as the instrument whereas the first difference in industry input tariffs interacted with firm skill share is served as the regressand. Again, the coefficient of industry input tariffs, the variable of our key interests, is negative and statistically significant. Thus, the 2SLS estimation results are consistent with our previous OLS estimates.

Table 5 2SLS estimates using panel data (2000–2006)

We now perform related statistical tests to check the validity of the instrument. The bottom module in Table 5 provides the first-stage estimates for all specifications. The coefficients of the instruments are negative and highly statistically significant, suggesting that it is more challenging to remove tariff barriers in industries with high initial tariffs. In addition, several tests were performed to verify the quality of the instruments. First, we use the Anderson canon correlated LM w² statistic to check whether the excluded instruments are correlated with the endogenous regressors. As shown in the upper module in Table 5, the null hypothesis that the model is under-identified is rejected at the 1% significance level. Second, the Cragg-Donald Wald F-statistic provides strong evidence for rejecting the null hypothesis that the first stage is weakly identified at a high significance level. The tests suggest that the instrument is valid and the specifications are well justified.

3.5 On the Possible Mechanism

The objective of this section is to discuss a possible mechanism to enrich our understanding of the main empirical finding that input trade liberalization leads to an increase in firms’ skill premium and provide some evidence for our theoretical conjecture. Inspired by the literature on “fair wages” (e.g. Egger & Kreickemeier, 2012), a possible mechanism to interpret our empirical findings is that skilled workers have greater bargaining power with their employers than unskilled workers.¹⁹ As a result, the incomes of skilled workers are more closely linked to firms’ profits but the incomes of unskilled workers are more in line with those of other firms in the same industry. Thus, a fall in input tariffs increases the firm’s economic profit, which in turn raises firms’ skill premium because the skilled labor commands a larger proportion of the incremental surplus than the unskilled labor.²⁰

To check whether such a conjecture is supported by the data, we replace the firm’s average wage, the regressand in our Mincer-type regressions, with the firm’s value-added per worker. Value-added per worker is one possible measure of labor productivity or, more generally, a proxy to the firm’s surplus per worker. If input tariff reductions raise the firm’s skill premium, we should observe that input trade liberalization also increases the firm’s value-added per worker because value-added per worker can be treated as another side of the same coin of firms’ skill premium. It is the core of our paper’s main hypothesis that intermediate input tariffs move value-added per worker in essentially the same way as they move the average wage (per worker), which is directly testable.

Specifically, we replace log of firm average wage with log of firm value-added per worker in the empirical specification in Eq. (7). To ensure that our estimation results are not contaminated by using the time-series proxy of the firm’s skill share, we focus on cross-section data in 2004 and report the estimation results in Table 5. The estimates in column (1) of Table 5 are obtained by using the ASIF customs matched data (as used in Table 3). After controlling a rich set of interacted region and industry fixed effects, the regression results show that the key coefficient of industry input tariffs interacted with skill share is negative and statistically significant, suggesting that input trade liberalization increases the firm’s value-added per worker.

The advantage of using ASIF-customs matched data is to allow us to govern firms’ importing status, but it is at the expense of reducing the number of observations since the matching between the two datasets (i.e., ASIF dataset and customs dataset) is imperfect (see more discussions in Yu, 2015). To see whether our findings are robust to different regression samples, column (3) runs the same regression as column (1) but instead uses the ASIF data set only. The key variable of interest, the interaction term between input tariffs and skill share, still exhibits a negative sign and statistically significant, indicating that our findings are robust by using different data sample.

Finally, we also replace the regressand of log average value-added with that of log per-worker profit and run the regressions using ASIF customs matched data in column (2) and sole ASIF data in column (4), respectively.²¹ Our key finding is robust in all specifications: When controlling for product-market tariffs in a firm’s industry, reduced input tariffs in a firm’s industry are associated with higher surplus per worker or overall profits at firms with skill intensive production (skilled workforces).

Although our interpretation is consistent with the evidence, it does not rule out other possible channels or mechanisms. There are other possible interpretations. For instance, an additionally employed skilled worker may generate a larger surplus, all else equal, and yet might receive a smaller share than unskilled workers (after bargaining). The large incremental surplus can be more than proportionally larger than the bargaining share difference to unskilled workers. Thus, skilled workers may seem to capture a larger proportion of the incremental surplus, but really they simply generate more surplus. However, we cannot validate this argument because it requires that the data contain variables that would directly measure the bargaining weight by skill groups, or related quantities.

4 Concluding Remarks

China has experienced dramatic tariff reductions since its accession to the WTO in 2001. On the other hand, wage inequality between skilled and unskilled labor of Chinese manufacturing firms has also increased significantly. To our knowledge, so far there is no study using micro-level evidence to explore the link between the two because there are no firm-level data on wages for skilled and unskilled labor. In this paper, we have developed a Mincer-type econometric approach to estimate firms’ skill premium based on imperfect Chinese firm level data on wage information. As in other ambitious attempts to investigate important issues with imperfect data, some compromises were made to conduct our estimations. Nevertheless, the finding that a fall in input tariffs is associated with an increase in the skill premium at firms with more skilled workforces is robust under different econometric specifications.

Such a finding is consistent with the idea that firms share the additional surplus generated by input trade liberalization mostly with skilled workers. Potential reasons for the observed increase in relative skill earnings at more skill intensive firms include technological and institutional factors: skills might be complementary with newly accessible foreign inputs on the technological side, or skilled workers might command stronger bargaining power over additional surplus generated under input trade liberalization.

Our findings also have rich policy implications. Trade liberalization can increase skill demand in China by prompting firms to use intermediate inputs that raise firms’ surplus. This happens either because less expensive or newly accessible inputs are complementary to skill or because skilled workers have a stronger bargaining power in their negotiation over the newly generated surplus. In any case, input trade liberalization is an appropriate policy instrument to foster firms’ surplus.