Wage inequality and trade reform: productivity channel

Wage inequality and trade reform:

productivity channel

Oleksandr Shepotylo

and Volodymyr Vakhitov

November 2, 2012

Abstract

We explore the productivity channel of rising wage inequality within a manufacturing industry. To solve the endogeneity problem, we use the trade and services liberalization episode in Ukraine in 2001-2007 as a source of exogenous variation. We conrm the main predictions of new models that link productivity and wages. First, rms that use liberalized goods and services more extensively have higher gains in productivity and and are more likely to export. These gains in turn result in heterogeneous growth in wages and higher wage inequality.

Second, extensive margins of trade non-linearly inuence industry in- equality of wages. The share of exporters within an industry increases

∗Kyiv School of Economics and Kyiv Economic Institute, oshepotylo@kse.org.ua

†Kyiv School of Economics and Kyiv Economic Institute, vakhitov@kse.org.ua

wage inequality when the share of exporters is low and reduces wage inequality when the share of exporters is close to 1.

1 Introduction

Rising wage inequality over the last 30 years, which according to OECD (2012) was the main contributor to the rising household income inequality, in both rich and poor countries is often linked to globalization. However, the evidence on the link is not clear cut. According to Goldberg & Pavcnik (2007), trade liberalization in Chile in 70's, in Mexico in 80's, in Argentina, Colombia, and India in 90's came along with increasing wage inequality.

OECD, on the other hand, reports that increased wage gap between bottom 10 and top 90 wage earners in the OECD countries came primarily through the skill-biased technical change and through the institutional and regulatory reforms.

The traditional Heckscher-Ohlin model fails to explain an increasing wage gap between skilled and unskilled labor that occured both in rich and poor countries over the last two decades. In addition, it cannot account for large variation of wages within a narrowly dened industry reported in the liter- ature (Dunne et al., 2004). Recently, Helpman et al. (2010, hereafter HIR) (and some others authors, including Egger & Kreickemeier (2012); Felber- mayr et al. (2011); Amiti & Davis (2012)) suggested a new class of models that link changes in wage inequality to the distributional changes in produc-

tivity within an industry. In their framework, heterogeneous rms, facing a labor market with search and matching frictions, in equilibrium pay wages that increase with productivity. In addition, exporting rms pay higher wages relative to rms with the same productivity in the closed economy.

As a result, distribution of wages within an industry has the same shape as the distribution of productivities. In the HIR trade openness inuences wage inequality within an industry through changes in the share of export- ing rms. The relationship is bell-shaped  trade openness increases wage inequality when trade openness is low and reduces wage inequality when trade openness is high.

This paper tests the main theoretical predictions of this class of mod- els, focusing mainly on the HIR model. First, the model predicts that the average wage across rms within an industry is an increasing function of productivity with constant elasticity of wage with respect to productivity.

Second, exporters pay higher wages, which jumps up discountiniously at the threshold productivity level separating exporters from non-expoters. Third, wage inequality within an industry is increasing when the share of exporting

rms is small and decreasing when the share of exporting rms approaches 1.

Testing these predictions is hard for several reasons. First, the wage- productivity relationship is plagued with endogeneity. In HIR, in equilibrium more productive rms employ workers with higher abilities. In addition, pro- ductivity within a rm depends on wages and bonuses that reward exerting

more eort.¹. Hence, the unobserved average ability and eort are positively correlated with productivity, causing an upward bias in the estimation of the eect of productivity on wages. Therefore, a source of exogenous vari- ation is needed to estimate the eect. We propose a well-known link from deregualtion to productivity as such variation. Recent studies of services and trade liberalization (Amiti & Konings, 2007; Arnold et al., 2011; Fer- nandes & Paunov, 2011; Khandelwal & Topalova, 2011) nd a positive eect of the liberalization on productivity of manufacturing rms. The size of the eect varies across rms because of dierences in intensity, with which

rms use liberalized goods and services as inputs. We use the episode of the Ukrainian trade and services liberalization in 2001-2007, isolated from other major deregulatory changes and driven by political pressure imposed by Ukraine's trading partners as a precondition for the Ukrainian WTO ac- cession, as the source of the variation. Shepotylo & Vakhitov (2012) nd that, on average, the episode led to a 9.2 percent increase in productivity of downstream manufacturing rms.

Second, price dierences within an industry are embodied in the out- put and productivity measures. Variation in prices may reect idiosyncratic demand shocks or market power, leading to overestimation of technical e- ciency of rmsFoster et al. (2008). It is not possible to separate the variation in productivity from the variation in markups directly, because we do not

1Lazear & Rosen (1981)on the eect of wage dispersion within a rm on the level of eort. Akerlof & Yellen (1990) on fair wages

observe rm-specic prices in the data. We introduce a demand system into the estimation of the production function, following De Loecker (2011), to eliminate the eect of the demand shocks. In addition, we control for the remaining variation in prices by computing the rm-specic markups follow- ing De Loecker & Warzynski (2009). As a result, we are able to mitigate the measurement error problem by disentangling the true productivity pass- through on wages from the markup pass-through.

Our main results are as follows. First, an increase in productivity leads to an increase in the average wage per worker, but the eect is not as strong as the OLS results suggest. The wage elasticity of productivity is 0.153 in the baseline specication. An increase in the markup also results in a higher wage, but the pass-through is twice lower in the baseline specication. Sec- ond, exporters pay a 6 percent wage premium. Wage elasticity of exporters is not statistically dierent from the wage elasticity of non-exporters, as the theoretical model suggests, which points that Pareto distribution of produc- tivity and wages ts reality quite well. Finally, we nd evidence supporting the bell-shaped relationship between the dispersion of wages and the share of exporting rms in the industry, but the results are not very robust statis- tically.

The rest of the paper proceeds as follows. Section 2 discusses a theoretical mechanism that links liberalization of services and distribution of wages in manufacturing industries. Section 3 describes the data. Section 4 outlines the identication strategy. Section 5 presents results. Section 6 concludes.

2 Fair wage hypothesis and its implications

2.1 Productivity and wages

In the Melitz model (Melitz, 2003), the productivity distribution of rms is

xed. The eect of trade liberalization on productivity comes only through the re-distribution of resources from less productive rms to more productive

rms. However, the literature documents a positive eect of trade (in Indone- sia and India - Amiti & Konings, 2007; Khandelwal & Topalova, 2011) and services (in the Czech Republic, Chile, and Ukraine  (Arnold et al., 2011), (Fernandes & Paunov, 2011), and (Shepotylo & Vakhitov, 2012)) deregu- lation on productivity of manufacturing rms. Hence, in addition to the redistribution eect, the industry level aggregates may change due to a shift in the whole distribution of productivities. Therefore, a more realistic model would recognize that as the economy opens up, distribution of productivity changes. In particular, trade liberalization have a direct, positive eect on productivity within a rm, with exporting rms beneting disproportionally more.

2.2 Theoretical framework

In HIR framework, workers ex-post are heterogeneous. Heterogeneous rms invest in screening to cut-o the least able workers. More productive rms have higher returns to screening, invest in screening to set a higher ability cut-o, and end up with workforce of higher average ability. Due to search

frictions and more costly replacement of higher ability workers, more produc- tive rms pay higher wages. When economy opens up, self-selection of rms into non-exporters and exporters raises the revenue of exporters relative to non-exporters, further inducing exporters to invest in screening. As a result, exporters pay higher wages for given productivity level. The resulting aver- age wage per worker ω as the function of productivity θ and export threshold θ_x is

ω = Υ(θ)^λω_d θ θd

η

(1)

where

Υ(θ) =











1, θ < θ_x Υ_x > 1, θ = θx

, θd is a zero prot productivity cut-o, and ωd is a wage set by the least productive rm. The functional form of this equation holds for production functions with either one or two factors of production, which allows us to employ two measures of productivity in the empirical analysis  labor pro- ductivity and total factor productivity  without changing the estimated equation. Figure 4 illustrates how the average wage per worker depends on productivity.

We extend HIR set up by allowing the productivity distribution to depend on the regulatory environment R.

G_θ(θ, R) = 1 − θ_min θ

z(R)

Figure 1: Wage and productivity

where _∂R^∂z < 0 and z > 2.² As a result of deregulation, attening of the productivity distribtion, represented by lower z in Figure 2, increases the coecient of variation as well as mean and variance of the disrtibution of productivities: ^∂E(θ)_∂R = ^−θ_(z−1)^min2

∂z

∂R > 0,^∂var(θ)_∂R = ^−2(z_(z−1)^2+z+1)θ3(z−2)^2min2

∂z

∂R > 0, and

∂CV (θ)

∂R = _[z(z−2)]^−z+11.5

∂z

∂R > 0.

Alternatively, regulatory reforms can increase the lower bound of pro- ductivities, θmin(Figure 3) , which would also lead to an increase in average productivity and higher variance :E(θ) = ^zθ_z−1^min,^∂E(θ)_∂R = _z−1^{z ∂θ}_∂R^min > 0. and var(θ) = _(z−1)^zθmin22(z−2),^∂var(θ)_∂R = _(z−1)^2zθ2^min(z−2)

∂θmin

∂R > 0, z > 2. However, it will

2Shepotylo & Vakhitov (2012) nd evidence that rms with higher productivity gain more from services liberalization, which can be modelled as the reduction in the shape parameter z.

Figure 2: Shift in distribution of rms due to regulatory changes. Increase in shape parameter

not change the coecient of variation, CV (θ) = √ ¹

z(z−2) and other scale in- variant measures of inequality within an industry. This result is crucially depend on the assumptions of Pareto distributed productivity. Cut-os in HIR are determined by the following equations

h Υ

δ kµ

x − 1i θ_x θ_d

_k^δη

= f_x

f_d (2)

and ˆ ∞

θmin

π(θ)dG(θ) = f_e (3)

where fd, fx, and fe are xed costs of producing, exporting, and entry.

k > 1 determines the shape of distribution of worker's ability and δ > 0 is a parameter of the screening technology. Consider a partial equlilibrium resut, when the reform does not eect aggregate demand. Clearly, the ratio ^θ_θ^x_d is

xed and does not depend on θmin. It can be also shown that the ratio _θmin^θd will not change as well. Therefore, the increase in minimum productivity will proportionally increase the productivity cut-os and will not change the inequality in wage distribution within an industry. The reduction in z, on the other hand, will have the eect on wage inequality.

As HIR show, for closed economy with rms' productivity distributed Pareto, distribution of wages within an industry is also distributed Pareto with minimum wage ωdand shape parameter 1 +_µ(z)¹ . In the closed economy µ(z) is the sucient statisics for wage inequality. Moreover, ^∂µ_∂z < 0, mean- ing that the regulatory reform that makes productivity distribution more

Figure 3: Shift in distribution of rms due to regulatory changes. Change in minimum productivity

Figure 4: Within sector inequality and trade openness

dispersed (lower z) leads to higher wage inequality. In an open economy, the wage inequality depends on µ and on extensive and intensive margins of trade. In particular, HIR show that the within-industry wage inequality increases when the share of exporters is small and decreases when the share of exporters is close to 1, as Figure 4 illustrates. Therefore, the regulatory reform in the open economy reinforces wage inequality relative to wage in- equality in the closed economy when the trade opennes is not very closed to 1.

3 Data

The data for the study come from several statistical statements annually submitted to the National Statistics Oce (Derzhkomstat) by all commer- cial rms in the country. The sample covers seven years from 2001 to 2007.

We start with the sample of manufacturing rms (NACE Section D) that never switched to another sector over the period of study. Since the Sectoral Expenditures Statement, required to construct the rm-specic service lib- eralization index, is submitted by only relatively large rms, our sample is restricted to the rms with 300 employees on average. We further excluded observations with zero or negative output, capital stock or employment as- suming that they indicated non-operational rms in a year. Based on the les accompanying the Enterprise Performance Statement and the Balance Sheet Statement, we have created a comprehensive prole for every rm which in- cludes the industry NACE), ownership type, exporting and importing status in every year.

As the measure of output, we used net sales after excise taxes from the Financial Results Statement. The Balance Sheet Statement is the source of the capital measure for which we used the end-of-year value of the tangible assets. For the production function estimation we used investments in tangi- ble assets which come from the Enterprise Performance Statement. The same statement is also a source for our employment variable. It is measured as the

year-averaged number of enlisted employees, which is a rough estimate of

the full time equivalent of labor used. The material costs come from the same statement in 2001-2004, whereas since 2005 they have been available from a separate Sectoral Expenditures Statement. The statement provides detailed information about the rm's expenditures on purchases from 22 manufac- turing sectors and 15 service sectors. Data from this statement were used to construct an individual rm-specic index of services liberalization and

rm-specic index of import tari liberalization. All variables were deated by the appropriate price deators as described in the next sub-section. The descriptive statistics for the sample are presented in Table 1.

Evolution of mean and standard deviation of productivity, markups, and wages in 2001-2007 is presented in Table 2. Over the investigated period, productivity has almost doubled, while dispersion of productivity has been moderately growing. Markups have also increased, but without noticeable trend in variation. Increase in markups during the trade and services liberal- ization episode is consistent with ndings of De Loecker et al. (2012)  during the trade liberalization episode in India, marginal costs have been reduced by 40 percent, while prices fall by 16.8 percent, leading to higher markups.

Average wage has increased by 221 percent, but wage dispersion has fallen.

The reduction in the wage dispersion goes against our prior expectations and against the predictions of the HIR model. Perhaps, the reduction in wage dispersion is related to privatization of state-owned enterprises, which set wages dierently from the private companies as we discuss in the robustness

Variable Observations Mean Std. deviation

ωit, thsd. UAH 2001 46515 4.885 3.94 Y_it, thsd. UAH 2001 46530 23521 180168

L_it, workers 46530 297.1 1360

Kit, thsd. UAH 2001 45923 8138 49820 M_it, thsd. UAH 2001 45924 15265 147687 I_it, thsd. UAH 2001 34321 2368 21711

ln(T F Pi,t) 44243 1.139 1.027

Importeri,t 46530 0.3131 0.4637

Exporteri,t 46530 0.3647 0.4813

Foreigni,t 46530 0.08764 0.2828

Exiti,t 46530 0.04008 0.1962

Entryi,t 46530 0.09654 0.2953

Urbani 46530 0.678 0.4673

Serv. Libit (EBRD) 46530 0.3966 0.6553 Serv. Libit (FDI) 46530 0.3784 0.7961

Privatei,t 46530 0.893 0.3092

Single plantit 46530 0.9356 0.2454

Input tariit 46530 5.693 3.2

Markupi,t 46515 5.468 153.8

Table 1: Descriptive statistics

Year ln(T F P

) ln(p

/c

) ln(ω

) Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.

2001 0.923 0.993 0.41 1.047 0.887 0.73 2002 0.995 0.99 0.361 1.057 1.135 0.66 2003 1.036 1.008 0.455 1.095 1.272 0.621 2004 1.177 1.003 0.535 1.064 1.403 0.587 2005 1.224 0.995 0.495 1.028 1.579 0.562 2006 1.364 1.057 0.696 1.049 1.751 0.549 2007 1.527 1.099 0.995 1.031 2.054 0.535 Total 1.139 1.027 0.516 1.066 1.371 0.695

Table 2: Productivity, markups, and wages

checks.

4 Empirical strategy

There are three features of the model that we test in the empirical analysis.

First, wages are positively linked to productivity with constant elasticity η that does not depend on export status. Second, exporters pay a positive wage premium. Third, wage inequality is higher within an industry where the share of exporters is small and is lower within an industry where the share of exporters is close to 1.

Our rst objective is to test whether higher productivity is associated with a higher average wage across rms in the same industry. We allow ωd

and θd to vary across regions and industries and change over time.³ Firm level productivity is unobservable to the researcher and measured with error due to unobservable prices. In addition, we allow wages inuence productiv- ity through eort level. Υx is determined by transportation cost and foreign market potential. We assume that transportation cost does not vary across industries and local market is growing at the same rate as the world econ- omy hence Υx does not vary across industries and over time. The empirical counterpart of the equation 1 in logarithmic form is

ln ω_it= α + η ln ˜θ_it+ µ × exporter_it+ X_itγ (4) + D_stµ + D_rr + _it

where ωit is rm i⁰s average wage at time t. ˜θit is rm i⁰s measured pro- ductivity at time t. exporterit is the dummy variable that takes the value of one if rm i exports in year t and zero otherwise. Dst are industry-year

xed eects, and Dr are region xed eects. X represents a vector of ad- ditional controls that inuence the reference wage ωd, including importer status, employment, whether the rm is located in urban area. There is also an important distinction between rms that exit at t+1 and surviving rms.

There is a lag between the decision to exit and the actual closing down. This

3ωd depends on the tightness of the labor market and on the xed costs of operating domestically, while θd depends on the xed costs of operating domestically. Since labor mobility in Ukraine is low, it is natural to assume that the tightness varies across regions.

Fixed costs are industry specic and can change over time with changes in the regulatory environment, R.

information is typically known by managers at time t, but is not observed by econometricians. Since we observe an exit ex post, we include a dummy for

rms, exiting next period.

4.1 Productivity measures

In the empirical analysis we use two measures of productivity  labor pro- ductivity and total factor productivity. Labor productivity is constructed in the straightforward manner as the value added deated by the industry price deator divided by the number of workers. To recover the TFP measure, we estimate the production function for each manufacturing industry (1-digit NACE classication) by the Olley-Pakes procedure (Olley & Pakes, 1996), controlling for sub-industry-specic demand and price shocks as suggested by De Loecker (2011). We identify demand and price shocks by exploiting variation in sub-industry (4-digit NACE classication) output at time t and by controlling for sub-industry and time xed eects. Under the constant elasticity of substitution (CES) demand system, unobserved prices are picked up by the variation in inputs and by aggregate demand and do not reect dierences in technology within an industry.

Technology and market structure

Consider a production technology of a single-product rm i at time t de- scribed by a production function

Y_it = L^α_it^lK_it^αkM_it^αmexp(˜ω_it+ ˜u_it), (5)

where Yit units of output are produced using Lit units of labor, Kit units of capital, and Mitunits of material and services inputs. ˜ωitis rm-specic pro- ductivity, unobservable by an econometrician, but known to the rm before it chooses a variable input Lit. ˜uit is an idiosyncratic shock to production that also captures measurement error introduced due to unobservable input and output prices.

Y_it is not known, because we do not observe rm-specic prices, pit. As a result observable sales, Rit = pitYit, may reect dierences in physical quantities and dierences in mark-ups across rms within the same industry.

Therefore, use of Rit as the dependent variable in estimation of produc- tion function parameters, without controlling for prices, determined among other things by market structure and demand shocks, would bias estimates of the production function if prices are correlated with inputs. Even more importantly, generated productivity estimates containing demand variation introduces a relationship between services liberalization and measured pro- ductivity through the impact of the liberalization on prices and demand.

To deal with this issue, we introduce a constant elasticity of substitution

demand system

Y_it= Y_st p_it P_st

σs

exp( ˜ξ_it), (6)

where Ystis total expenditures on goods produced by manufacturing industry s, in which rm i operates. Pst is industry-wide price at time t. ˜ξitis demand shock which is not observed by the rm when it chooses variable inputs in production, but determines dierences in mark-ups and measured produc- tivity. Assuming monopolistic competition, this demand structure implies a constant mark-up price-setting rule, which depends on the industry-specic elasticity of substitution σs. It further implies the following expression for the revenue function

Rit = (Yit)^σs+1σs (Yst)^−σs1 Pst



exp( ˜ξit)

−¹

σs . (7)

Substituting (5) into (7), moving Pst to the left-hand-side and taking logs yields

r_it = β_ll_it+ β_kk_it+ β_mm_it+ β_sy_st + ω_it+ ξ_it+ u_it, (8) where rit = ln(R_it/P_st)is log of revenue deated by corresponding industry (2 digit NACE) price deator, which is provided by the national statistics oce, and other lower-case letters represent upper-case variables in the log form.

β_f = ^σs_σ⁺¹

s α_f, where f = {l, k, m}. The elasticity of substitution in industry s can be retrieved as σs = −1/β_s. Finally, ωit = ^σs_σ⁺¹

s ω˜_it, ξit = −_σ¹

s

ξ˜_it, and u_it = ^σs_σ⁺¹

s u˜_it are error terms.

This procedure allows us estimating parameters of production function consistently, without having information on rm-specic prices. Here we rely only on the assumption that the demand function is of CES form, which is common in the literature.

Estimation of production function We estimate

r_it = β_ll_it+ β_kk_it+ β_mm_it+ β_sy_gt+ ω_it+ ξ_it+ u_it, (9) separately, for 11 manufacturing sectors (1 digit NACE classication). In what follows we suppress sector index for clarity of presentation. Capital and materials are deated by production price index. Instead of using over- all output of industry, we use more disaggregated sub-industry g output (NACE 4 digit), ygt, to add more variability to estimation of σs. It is valid since we assume that the elasticity of substitution is constant within the manufacturing sector.

We decompose the overall demand shock into the following components

ξ_it = ξ_t+ ξ_g+ ˜ξ_it, (10)

where ξt is industry-specic shock common to all rms at time t, ξg is de- mand factor aecting only rms producing in sub-industry g, and ˜ξit is an

Industry ln(K) ln(L) ln(M ) ln(Y ) Firms N

β_K α_K β_L α_L β_M α_M β_s

Food and Tobacco 0.032 0.033 0.199*** 0.207 0.737*** 0.766 0.038 2685 12165

(0.018) (0.016) (0.013) (0.026)

Textile and Leather 0.089* 0.090 0.426*** 0.430 0.482*** 0.487 0.010 884 3569

(0.036) (0.024) (0.018) (0.046)

Wood and Paper 0.032 0.040 0.172*** 0.217 0.703*** 0.885 0.206*** 571 2357

(0.023) (0.023) (0.037) (0.062)

Printing 0.061 0.061 0.407*** 0.404 0.511*** 0.507 -0.008 936 3939

(0.033) (0.033) (0.022) (0.043)

Coke, chemistry 0.108** 0.120 0.157*** 0.175 0.683*** 0.761 0.102*** 843 3967

plastics (0.036) (0.018) (0.032) (0.026)

Non-metallic minerals 0.036 0.037 0.151*** 0.156 0.788*** 0.812 0.029 795 3585

(0.029) (0.021) (0.019) (0.033)

Metallurgy 0.097** 0.105 0.173*** 0.186 0.676*** 0.728 0.072* 779 3243

(0.030) (0.027) (0.038) (0.032)

Machinery and 0.031 0.031 0.363*** 0.365 0.567*** 0.570 0.006 1085 4684

equipment (0.018) (0.025) (0.024) (0.032)

High-tech machinery 0.065 0.070 0.225*** 0.241 0.591*** 0.633 0.067 756 3531

(0.035) (0.027) (0.021) (0.051)

Vehicles and transport -0.041 -0.047 0.254*** 0.294 0.562*** 0.650 0.136* 321 1483

(0.042) (0.047) (0.041) (0.059)

Furniture and others 0.078 0.079 0.329*** 0.334 0.548*** 0.557 0.016 604 2458

(0.044) (0.046) (0.045) (0.083)

Notes: * p<0.05, ** p<0.01, *** p<0.001. Bootstrap standard errors are presented in parentheses. Table reports point estimates of revenue function parameters, β and production function paramters α = _σs^σ₊₁^s β, where σ^s = −1/βs for Ukrainian manufactruing rms for 2001-2007. Each row in the table represents Olley-Pakes estimation of production function for eleven manufacturing industries, dened according to NACE classication. Each estimation is performed with year and sub-industry dummies, which are not reported for brevity.

Table 3: Estimation of production function by Olley-Pakes procedure

idiosyncratic shock. Plugging in (10) in (9), we have the following equation

r_it = β_ll_it+ β_kk_it+ β_mm_it+ βy_gt+ δ_tD_t+ δ_gD_g+ ω_it+ ε_it (11)

where Dt is a a year xed eect and Dg is a sub-industry xed-eect. εit = ξ˜_it+u_itis the error term which is not correlated with inputs and productivity.

We estimate (11) by the Olley-Pakes methodology. Results are presented in Table 3. Total factor productivity net of price and demand eects is

recovered as

ln(T F P_it) = (r_it− β_ll_it− β_kk_it− β_mm_it− β_sy_st) σs

σ_s+ 1. (12)

4.2 Endogeneity of productivity measure

In the HIR framework, productivity is split into two components  techni- cal eciency θit and average ability of workers, ¯a, which is increasing with screening intensity. Since we do not have information on ¯a, our productivity includes both components. Since unobserved average ability is positively cor- related with both wage and productivity, it means that corr(¯a, ) > 0, and leads to an upward bias in estimation of η. Alternatively, in the framework of Egger & Kreickemeier (2012) and (Amiti & Davis, 2012) based on the fair wage hypothesis (Akerlof & Yellen, 1990), a wage rate below some level ω^∗, considered as a fair wage by workers, would lower their level of eort, e, lead- ing to lower productivity. The positive correlation between an unobserved eort and productivity, corr(e, θ) > 0, leads to a positive bias in the OLS estimation of the coecient η.

To estimate equation 4 consistently, a source of exogenous variation in productivity is needed. We propose a well-known link from deregulation to productivity as such variation. Recent studies of services and trade liberal- ization Amiti & Konings (2007); Arnold et al. (2011); Fernandes & Paunov (2011); Khandelwal & Topalova (2011) nd positive eect of the liberaliza- tion on productivity of manufacturing rms. The size of the eect varies

across rms because of dierences in intensity, with which rms use liberal- ized goods and services as inputs.

We use the episode of the Ukrainian trade and services liberalization in 2001-2007, isolated from other major deregulatory changes and driven by political pressure imposed by Ukraine's trading partners as a precondition for the Ukrainian WTO accession. Concerning services, the government de- veloped new laws and amended existing ones that regulated activities of TV and broadcasting, information agencies, banks and banking activities, in- surance, telecommunications, and business services. It led to dierentiated but positive eect on productivity of downstream manufacturing rms (She- potylo & Vakhitov, 2012). The results indicate that a standard deviation increase in services liberalization is associated with a 9.2 percent increase in TFP. In parallel with the services liberalization, the WTO negotiations also led to further liberalization of trade in i. To capture this eect in the empirical analysis we control for the eect of exporting on productivity and interact it with the eect of services liberalization to control for potential complementarities. In addition, we control for the positive spillovers from the liberalization of trade in inputs on productivity found, among others by Amiti & Konings (2007).

The index of services liberalization is rm-specic, reecting the variation in rm-level intensity of usage of various services inputs. Similarly to Arnold et al. (2011), but using rm level data, the index is computed according to

the following formula

serv lib_it =X

j

a^j_it× index^j_t (13)

where a^jit is the share of input sourced from the services sub-sector j in the total input for a rm i at time t, and index^jt is the measure of liberalization in the service sub-sector j at time t. The constructed index of services liberalization is further divided by a standard deviation for normalization.

We proxy for index^jt by EBRD indices of services liberalization and by the share of output produced by the rms with foreign ownership, which gives us two instruments.

The third instrument captures the rm-specic measure of trade liberal- ization. We compute an index of input tari liberalization and interact it with import status, following Amiti & Konings (2007):

input tarif f_it =X

s

b^s_it× tarif f_t^s (14)

where input tariffit is the rm-specic input tari, b^s_it is the share of input sourced from the two-digit NACE industry s in the total input for rm i at time t, and tarifft^s is the trade-weighted average MFN import tari in industry s at time t.

Our identication strategy would fail if we do not account for a general equilibrium eect of services and trade liberalization on the labor market. We

control for this by including industry-time specic eects, capturing struc- tural changes in the economy. The identication strategy would fail only if liberalization has systematic and dierentiated impact on wages within the industry that comes through channels other than the productivity channel.

4.3 Measurement error and omitted variable bias

The measured TFP suers from several shortcomings, including measurement errors due to imprecise input deators, variability of mark-ups across rms located within the same industry and variation in quality which is impossible to measure directly in our sample (Van Beveren, 2012). Consider for example, the measurement error due to the price variability. It introduces a bias in the measured productivity proportional to pit− p_st, where pit is the log of rm- specic price and pstis the industry price deator. Since corr(yit, p_it−p_st) < 0 and corr(yit, l_it) > 0, it introduces a negative correlation between labor and the error term, leading to a downward bias in the estimation of βl (as well as biased estimation of βk and βm). As a result, the measured productivity overestimate the technological eciency of a rm. Price variation within the industry may also reect dierences in quality across rms, producing similar goods. Depending whether dierences in prices fully capture variation in quality or not, corr(yit, quality_it) can be positive or negative, leading to the measurement error, but the direction of the bias is undetermined.

To control for unobserved price variation, we recover the rm-specic markups from the rm-level data following, the procedure developed by

De Loecker & Warzynski (2009). The markups are computed as

pit

c(θ_it) = βl

ω_itL_it/p_itY_it (15) where c(θit) is the marginal cost.

It should be noted that the constructed markup is endogenously deter- mined as a function of wages. Since corr(^p_θit^it, _it) < 0, the OLS estimation of 4 would lead to the downward bias in the estimation of the markup coecient.

Fortunately, we can instrument the markups by the same instruments as the productivity. For example, De Loecker et al. (2012) nd that after trade liberalization rms do not fully adjust prices to reduction in the marginal costs. The price declines are small relative to the declines in marginal costs, leading to a positive correlation between markups and productivity.

5 Results

5.1 OLS

We rst estimate equation 4 by OLS with industry-time and regional xed eects. The errors are corrected for heteroskedasticity at the rm level. The results are presented in Table 4. First, we regress ln ωit on the measured labor productivity and other controls, ignoring variation in the markups.

Column 1 of Table 4 shows that a 10 percent increase in labor productivity is associated with 2.4 percent increase in wage. Once we include markups

in column 2, the coecient on labor productivity is almost doubled, while the coecient on markups is negative and signicant. Comparing results in column 1 and 2, points to the omitted variable bias in column 1. At the same time, column 2 has two endogenous variables  productivity and markup  causing an upward bias in the OLS estimation of the coecient on productivity, while the estimation of the coecient on markup is biased downward. Inclusion of rm-specic eects in column 3 of the table alleviate endogeneity, resulting in the coecient on labor productivity almost the same as in column 1, but the coecient on the markup remains negative.

Column 3 shows that rms that start exporting pay a wage premium of 3.4 percent, which is well in agreement with the theoretical predictions. Adding the interaction between export status and productivity, exporterit× ln θ_it, in column 4 demonstrates that elasticity of wage with respect to productivity is not statistically dierent for exporters and non-exporters, which is also corresponds well with the model predictions. At the same time the coecient on export jumps to 0.08, indicating much higher export premium.

Results in column 4 also indicate that rms that start importing some of their inputs pay 4.8 percent more, which corresponds well with Amiti & Davis (2012), who found 3.2 percent increase in wage after the rm starts importing in Indonesia. Firms that switch ownership from domestic to foreign, dened as foreign ownership greater than of equal to 10 percent, pay 3.8 percent higher wages. Exiting rms pay 9.5 percent lower wage. We have not found any scale eect, measured by total employment, in the model that includes

rm xed eects, but when we compare rms of dierent sizes in columns 1 and 2, larger rms pay higher wages even after controlling for productivity and export status.

Columns 4 - 8 of Table 4 show results of the regression of ln ωit on TFP and other controls. The results with two dierent productivity measures are quite similar for exogenous controls. However, the positive eect of TFP on wages is smaller and the negative eect of markup on wages is much smaller in the absolute value. These results are expected because the model with labor productivity overestimates productivity for rms with high value of capital, leading to positive correlation of errors with productivity and negative correlation of errors with markups.

5.2 IV results

In this section we report results of estimation of 4 by IV, where we use three instruments  services liberalization measured by EBRD, services liberaliza- tion measured by the share of output produced by foreign-owned services companies, and the input tari liberalization measure, computed according to equations 13 and 14. The errors are corrected for heteroskedasticity at the

rm level, all regressions include industry-time and regional xed eects.

The results are presented in Table 5. We regress ln ωit on the measured TFP, markups, and other controls. Column 1 of Table 5 shows that a 10 percent increase in productivity is associated with 1.5 percent increase in wage, while a 10 percent increase in markup leads to 0.8 percent increase

Dependent variable:

ln(wagei,t) ln(vai,t/Li,t) ln(T F Pi,t)

OLS Markups FE Exp × OLS Markups FE Exp ×

prod. prod.

(1) (2) (3) (4) (5) (6) (7) (8)

Productivity 0.240*** 0.457*** 0.247*** 0.252*** 0.264*** 0.351*** 0.223*** 0.229***

(0.008) (0.022) (0.024) (0.029) (0.009) (0.011) (0.011) (0.012)

Markup -0.384*** -0.323*** -0.323*** -0.119*** -0.175*** -0.175***

(0.017) (0.021) (0.022) (0.006) (0.009) (0.009)

Exporteri,t 0.025** 0.036*** 0.034*** 0.080* 0.061*** 0.073*** 0.048*** 0.071***

(0.009) (0.007) (0.006) (0.035) (0.010) (0.009) (0.006) (0.012)

Exporter^i,t× -0.017 -0.020*

productivityi,t (0.013) (0.008)

Importeri,t 0.020* 0.080*** 0.048*** 0.048*** 0.114*** 0.155*** 0.060*** 0.061***

(0.010) (0.008) (0.005) (0.005) (0.010) (0.010) (0.006) (0.006) Foreigni,t 0.156*** 0.128*** 0.037** 0.038** 0.219*** 0.226*** 0.051*** 0.051***

(0.017) (0.014) (0.013) (0.013) (0.019) (0.018) (0.015) (0.015) ln(Li,t) 0.115*** 0.072*** -0.006 -0.006 0.106*** 0.091*** -0.024** -0.025**

(0.004) (0.003) (0.007) (0.007) (0.004) (0.004) (0.008) (0.008) Exiti,t -0.063*** -0.121*** -0.096*** -0.095*** -0.143*** -0.193*** -0.128*** -0.128***

(0.019) (0.016) (0.015) (0.015) (0.021) (0.021) (0.018) (0.018)

Urbani 0.094*** 0.064*** 0.015 0.014 0.132*** 0.132*** 0.025 0.024

(0.009) (0.008) (0.020) (0.020) (0.011) (0.011) (0.025) (0.026)

Industry × Year FE Yes Yes Yes Yes Yes Yes Yes Yes

Region FE Yes Yes Yes Yes Yes Yes Yes Yes

Firm FE No No Yes Yes No No Yes Yes

Observations 43985 43985 43985 43985 44236 44236 44236 44236

R² 0.583 0.715 0.601 0.602 0.509 0.529 0.522 0.522

Standard errors clustered by rms in parentheses. * p<0.05, ** p<0.01, *** p<0.001 Standard errors in parentheses

Table 4: OLS

in wage. Comparing this results with OLS results in column 5 of Table 4 conrms our theoretical priors that the OLS estimation of the coecient on productivity is biased upward and OLS estimation of the coecient on markup is biased downward. The coecient on productivity is more than halved, while the coecient on markup ips the sign, becoming positive and signicant.

Exporters pay a 6.1 wage premium relative to non-exporters, in accor- dance to the wage setting rule 1. Firms that import their inputs pay 9.6 percent more relative to rms that do not import. Foreign-owned rms pay 22 percent more relative to domestic rms. Elasticity of wage with respect to employment is positive and signicant 0.11, even after controlling for pro- ductivity, export and import status, and ownership. Exiting rms pay 11 percent less than surviving rms. In urban areas rms pay extra 13.8 per- cent, probably to compensate for the higher cost of leaving relative to the rural areas.

One of the concerns with the specication in column (1) is that wage is often set in advance and responds slowly to changing economic environment.

Column 2 reports results of the model specication

ln ω_it= α + η ln ˜θ_it−1+ µexporter_it−1+ X_it−1γ + D_stµ + D_rr + _it

which remarkably are very close to the results in column 1.

In column 3 we include an interaction term exporterit× ln θ_it. Elasticity

of wage with respect to productivity is not statistically dierent for exporters and non-exporters, which is also corresponds well with the model predictions.

At the same time the coecient on export becomes insignicant, indicating that we cannot discriminate between the models where wage of exporting

rms has a discrete jump or a dierent slope.

It should be noted that private rms in Ukraine pay part of the salary in cash and do not report it as the wage bill to evade the social security tax ranging from 32.6 to 49.7 percent of the wage bill. State-owned companies do not have this incentive, reporting all labor-related expenses in wage bills.

This feature of the tax system leads to under-reporting of wages by the private companies and can bias our results. Column 4 report the results with additional variable that control for state- vs. private-owned rms. As expected, private rms pay 16 percent lower wage, but it does not change our main conclusions.

Literature emphasizes distinction between multi-plant and single-plant

rms. Related to this, distinction between multi-product and single-product

rms is important when modeling rm's reaction to trade liberalization Bernard et al. (2011). We control for single- vs. multi-plant rms in column 5, which has no impact on our conclusions. In column 6 we control for new rms, which pay 8.5 percent lower wage than incumbents.

Our identication strategy would fail if exclusion restrictions are not valid.

For instance, trade and services liberalization can have a general equilibrium eect on wages which inuences more rms that use imported goods and

liberalized services more intensively. In that case our excluded variables inuence wages not only through productivity, but also directly. We are quite condent that our estimation is valid for several reasons. First, we control for the industry-time specic trends directly. Second, the overindentication test does not reject validity of our instruments. Finally, when we include additional controls for input taris at sub-industry level. It does not change our results as column 7 indicates.

When we include all variables that we use in the robustness checks, it only strengthens our results. Column 8 of Table 5 shows that 10 percent increase in productivity leads to 1.7 percent higher wage, while 10 percent increase in the markup leads to 1.37 percent increase in wage. Coecients on exporterit

and exporterit × ln(θ_it) are jointly signicant, but we can not distinguish which model  with a jump or with change in slope  better represents the data.

To sum up, we nd a robust positive eect of productivity on wages, with elasticity in the range 0.14-0.17. The markups also positively contribute to wages, but the eect is weaker. Exporting rms pay a wage premium of 3-6 percent, depending on the model specication. The result is not robust to inclusion of an interaction term exporterit × ln(θ_it), meaning that we can not distinguish between the model with a dierent intersection for exporters vs the model with a dierent intersection and dierent slope. Overall, our results conrm the predictions of HIR at the rm level.

Dependent variable:

ln(wageit) (1) (2) (3) (4) (5) (6) (7) (8)

Base Lagged Inter- Private Single Entry Input All

action plant tari

ln(T F P_i,t) 0.153*** 0.155*** 0.142*** 0.165*** 0.156*** 0.157*** 0.145*** 0.167***

(0.033) (0.037) (0.041) (0.033) (0.033) (0.033) (0.034) (0.044) Markupi,t 0.083*** 0.074*** 0.083*** 0.089*** 0.083*** 0.085*** 0.140*** 0.137***

(0.014) (0.016) (0.014) (0.014) (0.014) (0.014) (0.024) (0.024) Exporteri,t 0.061*** 0.050*** 0.032 0.056*** 0.060*** 0.060*** 0.047*** 0.042

(0.012) (0.013) (0.022) (0.012) (0.012) (0.012) (0.013) (0.023) Importeri,t 0.096*** 0.089*** 0.094*** 0.102*** 0.096*** 0.097*** 0.073*** 0.085***

(0.016) (0.017) (0.015) (0.016) (0.016) (0.016) (0.018) (0.016) Foreigni,t 0.220*** 0.220*** 0.220*** 0.232*** 0.221*** 0.221*** 0.211*** 0.228***

(0.020) (0.022) (0.020) (0.020) (0.020) (0.020) (0.021) (0.020) ln(L_i,t) 0.115*** 0.119*** 0.115*** 0.111*** 0.118*** 0.110*** 0.122*** 0.115***

(0.005) (0.006) (0.005) (0.005) (0.005) (0.005) (0.006) (0.005) Exiti,t -0.117*** -0.119*** -0.115*** -0.115*** -0.119*** -0.096*** -0.099***

(0.024) (0.025) (0.024) (0.024) (0.024) (0.026) (0.026) Urbani 0.138*** 0.138*** 0.138*** 0.141*** 0.138*** 0.140*** 0.137*** 0.142***

(0.012) (0.013) (0.012) (0.012) (0.012) (0.012) (0.013) (0.012)

Exporteri,t× 0.026 0.001

ln(T F Pi,t) (0.023) (0.025)

Privatei,t -0.164*** -0.176***

(0.017) (0.017)

Single planti,t 0.073*** 0.091***

(0.018) (0.019)

Entryi,t -0.085*** -0.096***

(0.013) (0.016)

Input tarij,t -0.009** -0.007*

(0.003) (0.003)

Industry × Year FE Yes Yes Yes Yes Yes Yes Yes Yes

Region FE Yes Yes Yes Yes Yes Yes Yes Yes

Observations 44236 33222 44236 44236 44236 44236 44236 44236

Hansen J statistic 0.055 1.263 0.047 0.000 0.044 0.058 0.049 0.001

χ²(1)p-value 0.815 0.261 0.829 0.999 0.834 0.810 0.824 0.978

R² 0.470 0.440 0.469 0.472 0.471 0.471 0.436 0.447

Standard errors clustered by rms are reported in parentheses. * p<0.05, ** p<0.01, *** p<0.001

Table 5: IV

Figure 5: Average wage and share of exporters

5.3 Industry level results

Having conrmed the positive causal eect from productivity to wages, we further test predictions of HIR at industry level. We aggregate our data to the level of NACE 3 digit sub-industries. We proxy ρ = θd/θx by the share of exporters, Njt^x/N_jt, where Njt^x is the number of exporters and Njt is the total number of rms in sub-industry j at time t. Figure 4 presents a scatterpolot of the average wage as a function of the share of exporters. The solid line is the quadratic t that minimizes the total sum of squared errors of the follow- ing optimization: minc^m,α^m,β^m

hP

j,t(ln ω_jt− c^m− α^mexpshare_jt− β^mexpshare²_jt)²i.

Clearly, the average wage increases with the share of exporters within the sub-industry.

Figure 6: Coecient of variation and trade openness

Figure 5 presents a scatterpolot of the coecient of variation, cv(ln ωjt) =

σ(ln ωjt)

E(ln ωjt), as a function of the share of exporters. The solid line is the quadratic

t that minimizes the total sum of squared errors of the following optimiza- tion: minc^cv,α^cv,β^cv

hP

j,t(cv(ln ω_jt) − c^sd − α^sdexpshare_jt− β^sdexpshare²_jt)²i.

As predicted by the HIR model, the coecient of variation of wages within a sub-industry is rst rising and then declining with the increase in trade openness.

The observed regularities might be driven by variation of wages across industries and by other factors. To test the predictions on the relationship between sub-industry wage statistics and trade openness, we estimate the

following equations

ln ω_jt = c^m+ η^mln θ_jt+ α^mexpshare_jt+ β^mexpshare²_jt+ X_jtγ^m+ u_jt

sd(ln ω)_jt = c^sd+ η^sdσ(ln θ)_jt+ α^sdexpshare_jt+ β^sdexpshare²_jt+ X_jtγ^sd+ v_jt

and

cv(ln ω)_jt = c^cv+ η^cvσ(ln θ)_jt+ α^cvexpshare_jt+ β^cvexpshare²_jt+ X_jtγ^cv+ v_jt

where sd(ln ω)jt and cv(ln ω)jt are standard deviation and coecient of vari- ation of wages within the sub-industry j at time t. We estimate these equa- tions with and without industry eects (industry is dened according to 1 digit NACE classication), to explore the relationship within and between industries. The results indicate that the average wage within sub-industry is positively linked to average productivity, while higher variation of wages within industry is positively linked to higher variation of productivity. The coecients on exporter shares have expected sign and is signicant. Overall, we can conclude that the results do not reject the hypothesis on inverse U- shaped relationship between trade openness in the industry and variation of wages.

Dependent variable Average wage^j,t St. Dev. Wage^j,t Coef. Var. Wage^j,t

(1) (2) (3) (4) (5) (6)

OLS FE OLS FE OLS FE

ln(T F Pj,t) 0.054*** 0.523***

(0.015) (0.054)

σ(ln(T F Pj,t)) 0.089* 0.101* 0.135** 0.128*

(0.040) (0.043) (0.050) (0.050) Share of exportersj,t -0.399 -0.147 0.657*** 0.694*** 0.460** 0.384*

(0.267) (0.280) (0.133) (0.148) (0.164) (0.186) Share of exporters²j,t 0.368 0.072 -0.831*** -0.856*** -0.607*** -0.536***

(0.252) (0.245) (0.121) (0.130) (0.137) (0.155) Share of importers^j,t 0.493*** 0.410*** -0.004 0.003 -0.230* -0.371***

(0.129) (0.121) (0.067) (0.068) (0.095) (0.100) Share of foreignj,t 0.659*** 0.481*** 0.242* 0.265** -0.033 -0.047 (0.161) (0.136) (0.107) (0.099) (0.083) (0.101)

ln(Lj,t) 0.061* 0.101*** -0.017 -0.010 0.010 0.063**

(0.027) (0.025) (0.012) (0.014) (0.015) (0.020) Share of exiting rms^j,t -1.954*** -1.389*** 0.265 0.277 0.808* 0.786*

(0.294) (0.258) (0.243) (0.246) (0.318) (0.309) Share of urbanj,t 0.346*** 0.302*** 0.035 0.101 -0.153** -0.154*

(0.084) (0.088) (0.038) (0.055) (0.058) (0.076)

Sub-Industry FE No Yes No Yes No Yes

Observations 683 683 681 681 681 681

R² 0.369 0.618 0.206 0.251 0.196 0.269

Standard errors clustered by industries in parentheses. * p<0.05, ** p<0.01, *** p<0.001

Table 6: Industry level results

6 Conclusions

We nd that the productivity channel plays an inportant role in the increase in wage inequality, observed since the beginning of 90's XXX

References

Akerlof, G. & Yellen, J. (1990). The fair wage-eort hypothesis and unem- ployment. The Quarterly Journal of Economics, 105(2), 255283.

Amiti, M. & Davis, D. (2012). Trade, rms, and wages: Theory and evidence.

The Review of Economic Studies, 79(1), 136.

Amiti, M. & Konings, J. (2007). Trade liberalization, intermediate inputs, and productivity: Evidence from indonesia. American Economic Review, 97(5), 1611  1638.

Arnold, J., Javorcik, B., & Mattoo, A. (2011). Does services liberalization benet manufacturing rms?: Evidence from the czech republic. Journal of International Economics, 85(1), 136146.

Bernard, A., Redding, S., & Schott, P. (2011). Multiproduct rms and trade liberalization. The Quarterly Journal of Economics, 126(3), 12711318.

De Loecker, J. (2011). Product dierentiation, multiproduct rms, and esti- mating the impact of trade liberalization on productivity. Econometrica,