Capital-skill complementarity, productivity and wages: Evidence from plant-level data for a developing country

Available online at www.sciencedirect.com

Labour Economics 15 (2008) 1 – 17 www.elsevier.com/locate/econbase

Capital-skill complementarity, productivity and wages: Evidence from plant-level data for a developing country Mahmut Yasar a,⁎, Catherine J. Morrison Paul b,1 a

b

Department of Economics, The University of Texas at Arlington, 701 S. West Street, Box 19479, Arlington, TX 76019, United States Department of Agricultural and Resource Economics and the Giannini Foundation, University of California, Davis, One Shields Avenue, Davis, California 95616, United States Received 31 January 2006; received in revised form 19 December 2006; accepted 4 January 2007 Available online 13 January 2007

Abstract Different types of labor and capital inputs have varying productive contributions that are dependent on plant characteristics. We estimate such contributions and their underlying determinants, recognizing the interactions among labor and capital components that reflect their substitutability or complementarity, for Turkish manufacturing plants. We distinguish technical and non-technical labor, and structures, machinery and computer capital, as well as the shares of female workers and imported capital in our production function specification. We find capital-skill complementary for both machinery and computers; greater productive contributions and thus wages for skilled labor are associated with more machinery intensity and computer use. The reverse is true for unskilled labor, which is complementary only with capital structures. Our results suggest that synergies among skilled (technical) labor, computers, and machinery capital have productivity- and skilled wage-enhancing effects that could contribute to productivity convergence of developing toward developed countries, even with their differing industry and input composition. © 2007 Elsevier B.V. All rights reserved. JEL classification: J24; J31; O14 Keywords: capital-skill complementarity; labor and capital composition; plant productivity; wage determinants

⁎ Corresponding author. Tel.: +1 817 272 3061; fax: +1 817 272 3145. E-mail addresses: [email protected] (M. Yasar), [email protected] (C.J. Morrison Paul). 1 Tel.: +1 530 752 0469; fax: +1 530 752 5614. 0927-5371/$ - see front matter © 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.labeco.2007.01.001

2

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

1. Introduction Questions about the productive contributions of and returns to different labor and capital components are often raised in the literature on firms' productivity determinants, and are typically addressed using aggregate data for developed countries. Some of these questions involve interactions between labor and capital types, such as the relationship between skilled labor and technology embodied in capital, often called the capital-skill complementarity hypothesis (Griliches, 1969). More general labor-technology linkages are alternatively represented as skillbiased technical change - usually disembodied technical change but sometimes related to specific technologies such as computers. Many studies in the labor economics literature address such issues by evaluating wage patterns - in particular, increasingly disparate wages for skilled as compared to unskilled workers as technology advances and trade expands. Studies in this literature are often based on estimation of a share equation for skilled labor, with relative skilled-unskilled labor wages or levels as determinants (Berman et al., 1994, 1998; Hollanders and ter Weel, 2002). Other typical arguments of the function include capital and output levels (or ratios), technology proxies (high-tech capital, computers, R&D, or disembodied technical change), and trade indicators (imports and exports). Some studies estimate a wage or demand equation for skilled labor rather than a share equation that captures both price and quantity variation (Revenga, 1992; Krueger, 1993; Bernard and Jensen, 1995; Van Reenen, 1996). Siegel (1997) similarly estimates a model based on a proxy for labor quality (price), and Hellerstein and Neumark (2004) imbed a labor quality expression in a (Cobb-Douglas) production function. In other cases both a production function and wage or share equation are estimated (Dupuy, 2003), although economic optimization is not necessarily used to derive the equation system (Hellerstein et al., 1999). Most studies in this literature identify a connection between the wages or productivity of skilled labor with a capital (or technology) intensive input mix, implying capital- or technologyskill complementarity.2 That is, increasing the capital- or technology-intensity of production appears both to enhance productivity directly, and to imply more skilled labor intensity and higher wages. Overall, the typical findings of this literature may be summarized as technological factors embodied in capital increasingly driving high demand for and returns to more skilled labor in developed countries.3 However, as emphasized by Goldin and Katz (1998), capital-skill relationships may differ by country, industry, and time period. In this study we evaluate input composition patterns for Turkish manufacturing plants, with a focus on capital-labor relationships that might be expected to differ from those in more developed countries. For example, such relationships might be different in a country such as Turkey because the plants rely more on imported capital (rather than developing their own technology) and the primary industries are less high-tech, or because both capital and skilled labor intensity are therefore lower. For our analysis we use an unbalanced panel of plant-level data for 886 manufacturing firms over three years (1995–1997; 2598 observations) and four industries (apparel, textiles, motor 2 Doms, Dunne and Troske (1997), Bresnahan, Brynjolfsson and Hitt (2002), and Leiponen (2005), for example, find skill-biased technical change or technology-skill complementarity associated with high-tech capital at the plant level, which affects labor composition, wages, labor productivity, and profits. 3 Although a large proportion of the studies in this lieterature are for the U.S., Betts (1997) and Caroli and Van Reenen (2001) find skill-biased technology change for Canadian manufacturing industries and British and French establishments, respectively, Haskel and Heden (1999) find complementarity between computers and skilled labor for establishments in the U.K., and Machin and Van Reenen (1998) find such patterns for seven OECD countries.

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

3

vehicles and parts, and meat processing). Together these industries account for a large fraction of Turkish manufacturing output. Textiles and apparel are sub-sectors of the textile, wearing apparel, and leather industry (ISIC 32), and the meat processing sector is a sub-category of the food, beverages and tobacco industry (ISIC 31). Along with the motor vehicles and parts industry, these industries account for approximately 50, 43, and 48 percent of Turkish manufacturing employment, output, and exports, and thus well reflect the differential industry mix in such a developing compared to a developed country. We distinguish two types of labor plus a gender indicator (numbers of administrative, engineering, and other technical versus non-technical labor, and female labor share), to consider their differential productivity contributions and drivers. We also recognize two types of capital plus high-tech and imported capital indicators (stocks of structures and transportation equipment versus machines, a dummy variable for computer use, and the imported capital share), to evaluate the impacts of capital intensity, computerization, and imported technology on productivity and wages. Our production function model represents the plants' technology, allows for interactions among the arguments of the function (flexibility), and includes fixed effects for industry, region, year, and size. This equation is estimated along with a share equation for overall labor, representing the choice of labor composition and its dependence on capital levels and mix. Decomposition of the share equation allows us to separately capture the wage (returns) and output effects underlying labor share and composition.4 Further, the flexible form of our production function allows us to explore in detail both the productive contributions of the labor components and their underlying determinants (capital and other factors). Output elasticities (proportional marginal products) representing the contributions of the two labor types as well as capital and materials inputs are measured as first order elasticities of the production function, as are the effects of the female labor share, computer use, and imported technology. Second order relationships or elasticities are in turn measured to capture inputspecific cross effects such as the impacts of greater capital intensity or computer use on the contribution and returns to skilled (technical) and unskilled (non-technical) labor. We find direct productivity impacts of computer use and capital machinery intensity that are augmented by higher levels of skilled labor through capital-skill complementary. Conversely, this relationship implies greater productive contributions and value (and thus wages) of skilled labor associated with computer use and higher capital machinery intensity. Reversed relationships are found for unskilled labor, which is complementary only with capital structures, although the wage effects are insignfiicant. This suggests that expanding capital intensity and particularly computerization could stimulate productivity and overall returns to labor in Turkish manufacturing through synergistic effects with skilled labor, contributing to “catch-up” with the EU and other developed countries. 2. The model and measures We wish to identify variations in productive contributions for different types of labor, and their dependence on capital levels and composition, for the Turkish manufacturing plants in our data. 4

Our model is related to Hellerstein et al. (1999), who estimate a production function with a wage equation that depends on the plant’s labor composition, although our labor share equation is explicitly derived from economic optimization and our functional form captures interactions among inputs. It is also similar to Hollanders and te Weel (2002) although they do not estimate the production function and they focus on R&D as their technological indicator.

4

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

Table 1 Variable definitions Variable

Definition

ln Y ln LT ln LN ln KM ln KS KC ln M ln E pL pY pKM pKS IS FS

Natural logarithm of deflated output Natural logarithm of skilled labor (administrative + engineers + technicians) Natural logarithm of unskilled labor (non-technical) Natural logarithm of deflated machinery capital input Natural logarithm of deflated structure and transportation capital input =1 if the plant invested in computers in year t and zero otherwise Natural logarithm of deflated material input Natural logarithm of deflated energy input Labor price (total compensation divided by the number of laborers) Output Price Machinery Capital Price Structure Capital Price Imported machines and equipment/total investment Share of female workers in total employment

Such patterns can be represented through a model of the technology that recognizes different labor components, and interactions of the labor types with other productive factors - in particular capital components. The technology may generally be represented by the transformation function T(Y,X, R, D) = 0, where Y is a vector of outputs, X is a vector of inputs, R is a vector of external factors or plantlevel characteristics, and D is a vector of dummy variables or fixed effects. By the implicit function theorem, if T (Y,X, R, D) is continuously differentiable and has non-zero first derivatives with respect to one of its arguments, T(Y, X, R, D) may be specified (in explicit form) with that argument on the left-hand-side of the equation and the other arguments on the right-hand-side. Thus, we can define the asymmetric transformation function Y1 = F(Y−1, X, R, D), where Y1 is a chosen numeraire, and Y−1 is the vector of all outputs except Y1, that represents the overall technology but does not imply economic choices. For a model with only one output, this function becomes the more commonly estimated production function Y = F(X, R, D). For our purposes, we further distinguish the inputs in this function as Y = F(K, L, M, R, D), where K, L and M are vectors of capital, labor, and materials components, respectively. For our application each of these vectors has two components. The K variables are KM = machines, and KS = structures and transportation equipment.5 Additional (high-tech and trade) capital indicators, a dummy variable indicating whether a plant reports any computer investment, KC,6 and an imported/domestic investment share, IS, are also included in the R vector. The L variables are the employment levels for LT = technical, administrative and engineering labor, and LN = nontechnical labor, which broadly represent a split between skilled and unskilled workers. The share of female workers, FS, is also included as a component of the R vector. The M variables are E = energy and M = intermediate materials. The D vector includes industry, year, and size7 dummy variables.

5 Structures (buildings) and transportation equipment were aggregated due to the existence of zeroes for the investment series for both types of capital. 6 We were unable to use these data to construct a computer stock variable due to the short time period of the data, which made identifying a benchmark stock level and cumulating the stock difficult, and many zeroes for investment levels which are difficult to accommodate with a logarithmic function. 7 Small plants are defined as those with less than 50 employees, medium are with 50–100 employees, and large plants are those with more than 100 employees.

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

5

Table 2 Descriptive statistics (Constant Value Quantities at 1987 Prices, in ‘000 Turkish Lira) Continuous Variables

Mean

Standard Deviation

Minimum

Maximum

Output (Y ) Skilled or Technical Labor (LT) Unskilled or Non-Technical Labor (LN) Machinery (KM) Structure and Transportation (KS) Material (M ) Energy (E ) Labor Price ( pL) Output Price ( pY) Machinery Price ( pKM) Structure Price ( pKS)

7,571.15 36.86 116.89 378.74 45.18 5,603.96 195.51 300.05 115.97 93.34 180.09

38,475.17 100.97 235.70 2,438.01 225.21 26,602.18 905.88 287.48 49.17 37.11 79.79

11.03 1.00 1.00 0.02 0.003 0.19 0.06 7.80 64.87 52.90 94.40

756,834.25 2522.00 4845.00 55,879.07 6,912.39 616,738.31 22,178.53 2,610.94 201.81 145.70 290.80

Dummy Variables and Shares

Mean (Percentage of plants for dummy variables)

KC IS FS Small Medium Large Meat Processing Industry (3111) Textile Industry (3212) Apparel Industry (3222) Motor Vehicles and Parts Industry (3843) South and East Anatolian Region Central Anatolian Region Blacksea Region Mediterranean Region Agean Region Marmara Region Total Number of Observations

31.22 5.80 33.88 39.72 22.75 37.53 7.35 11.32 58.82 22.52 2.73 7.81 1.27 4.08 19.71 64.40 2598

Table 1 summarizes these variable definitions, and Table 2 provides summary statistics (the data construction is detailed further in the Appendix). Alternative functional forms may be used to approximate the F(•) equation for estimation. The most common are the Cobb-Douglas (CD) and translog (TL) forms, which are first-and secondorder logarithmic functions. Their logarithmic forms imply that production relationships, as well as economic optimization, are most naturally represented in terms of output elasticities. That is, the output elasticities (productive contributions or “shares”) εY,j = ∂ln Y/∂ln Xj = MPj Xj/Y, where MPj = ∂Y/∂Xj is the marginal product of the jth input, are the primary first order measures typically estimated for analysis of production relationships. Further, the sum of these elasticities, εY,X = ∑j εY,j, represents returns to scale (εY,X N 1 implies increasing returns). The εY,j are functions of X,R only if F(•) is approximated by a second order or flexible functional form (such as the TL); εY,j is a parameter for the CD form.8 Fixed effects in D, or other measures without second order (interaction or cross) terms included, will not appear as arguments

8

For the CD it will simply be a parameter and therefore estimation of the function is moot.

6

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

of these elasticities, so their contributions to input mix may not be addressed. Even if interaction terms are omitted, however, the direct productive impacts of shift variables may be measured by output elasticities analogous to those for the inputs: εY,r = ∂ln Y/∂Rr and εY,d = ∂ln Y/∂Dd for arguments in the R and D vectors. If profit-maximization (optimization) at the plant level is assumed, the choice of input j is implied by the equality pj = pY MPj = pY ∂Y/∂Xj, where pj is the price of input j and pY is the price of output. The output elasticities can then explicitly be expressed as input share equations, which may be used for estimation with the production function: Sj ¼ pj Xj =pY Y ¼ AY =AXj ðXj =Y Þ ¼ AlnY =Aln Xj ¼ eY ; j ðX ; RÞ;

ð1Þ

where Sj is the revenue share of input j.9 Such equations represent relative input demands rather than input demand levels. That is, the change in input j's share with a change in an argument of F(•) captures both the input-specific impact on the marginal product (and implicitly use) of Xj and the overall productive impact. It thus reflects the impact on input Xj's value of a change in, say, Xk, relative to the impacts on overall inputs and thus output, which is interpreted as an input “bias.” More formally, given the definition Sj = MPj Xj/Y and using the quotient rule, the bias in input j use with a change in Xk is: Bj;k ¼ ASj =AlnXk ¼ ðASj =AXj Þ&Xk ¼ Sj ðAlnMPj =AlnXk −AlnY =AlnXk Þ ¼ Sj ðej;k −eY ;k Þ; ð2Þ where εj,k = ∂lnMPj/∂lnXk is the elasticity of (the marginal product of) input j with respect to the use of input k. For example, for labor input Xj and capital input Xk, Bj,k reflects the substitutability or complementarity of the labor and capital components. If additional Xk increases the Xj share (Bj,k N 0) the positive impact on the marginal product of Xj exceeds the overall productivity effect, implying Xj−Xk complementarity. Note also that the share equations embody information on input prices or wages, implicitly capturing the choice of input composition (or quality). That is, for labor component Xj, given that Sj = pjXj/pYY = ∂lnY/∂lnXj, pj ¼ AlnY =AlnXj &pY Y =Xj ¼ Sj &pY Y =Xj :

ð3Þ

This expression therefore represents the dependence of the Xj price on Y and Sj, which are in turn dependent on all the arguments of the production function. It can thus broadly be interpreted as a hedonic price equation relating the wages in a particular plant to its labor composition, as well as to other input (i.e., capital) levels and plant characteristics. The implied difference in the price (wage) of labor input Xj at different levels of another input Xk, expressed in the elasticity form ∂ln pj/∂ln Xk, is therefore: epj; Xk ¼ Alnpj =Aln Xk ¼ pY =pj Xj &ðASj =Aln Xk &Y þ AY =Aln Xk &Sj Þ ¼ pY Y =pj Xj &ðASj =AlnXk þ Sk Sj Þ ¼ ðBj;k þ Sj Sk Þ=Sj :

9

ð4Þ

This treatment is similar to Dupuy (2003), although the use of cost instead of revenue shares in that study implies that the revenue and cost shares are equivalent, in turn implying constant returns to scale.

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

7

Differences in wages across plants can thus indirectly be attributed to the various arguments of the production function through the bias measures Bj,k, but the full impact on pj depends also on the shares of Xj and Xk. In sum, we wish to compute measures to evaluate the productive impacts of different labor types, and their interactions with other inputs, that provide insights about capital-skill complementarity and other output and input composition relationships, and the resulting implications about productivity and wages. These measures are represented by (first order) output elasticities such as εYj, and (second order) bias and input price measures such as Bj,k and εpj,Xk, which provide a wide range of implications about plants' productive patterns. 3. Empirical implementation To move toward empirical implementation of our model we must specify a flexible functional form that captures second order relationships representing the impacts of, for example, a particular type of capital on the “productivity” of a particular type of labor. We also wish to allow for the possibility that, as suggested by Dupuy and de Grip (2006),10 computer use might not only directly affect production but also interact with input substitution parameters. We thus use a translog form for estimation, with the addition of third order interaction terms to allow for impacts of computer use on input interactions. Such a form can be written for the general production function Y = F(X,R,D) for Turkish manufacturing plants as: lnYit ¼ a0 þ Rd bd Dd þ Rj bj ln Xjit þ Rr br Rrit þ Rr Rj grj Rrit ln Xjit þ :5Rj Rk djk ln Xjit ln Xkit þ :5Rj Rk djkc KC ln Xjit ln Xkit þ lit1 ;

ð5Þ

where i and t are plant and time subscripts (hereafter suppressed for notational simplicity), j,k denote K, L, and M variables in the X vector, r,d denote variables in the R and D vectors, and μit1 is assumed to be a random error. Such a function allows us to evaluate the significance of both the third order computer interaction terms (to test the Dupuy and de Grip, 2006, hypothesis), and the second order terms (to test the validity of separability assumptions or even a first order CobbDouglas model, as well as the significance of cross terms for the R variables). For estimation, this function is augmented by a share equation for overall labor: SL ¼ SLT þ SLN ¼ AlnY =Aln LT þ AlnY =Aln LN ¼ eY ;LT þ eY ;LN ¼ Rm bm þ Rr Rm grm Rrit þ :5dTN ðln LTit þ ln LNit Þ þ Rm dmm ln Lmit ; þ :5Rm Rk p m djk ln Lmit ln Xkit þ :5Rm Rk p m djkC KC ln Lmit ln Xkit þ lit2 ; ð6Þ where m,n denote the LT and LN labor components and SL is measured as the ratio of the wage bill to sales value. Although we estimate one labor share equation because we do not have separate wage data for technical and non-technical labor, the disaggregation of the labor input allows us to interpret the parameter estimates of F(•) as contributions to the share and price of each labor type, as well as to overall labor.

10

This possibility was suggested by an anonymous referee.

8

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

That is, the “decomposition” of the share equation into its input-specific (marginal product) and output components (Eq. (2)), and the specification of the input price “hedonic” expression (Eq. (4)), can be derived from this overall labor share equation. Given that SL = SLT + SLN, ASL =Aln Xk ¼ ASLT =Aln Xk þ ASLN =Aln Xk ¼ BLTk þ BLNk ¼ SLT ðeLT;k −eY ;k Þ þ SLN ðeLN ;k −eY ;k Þ ¼ SLT ðeLT ;k −Sk Þ þ SLN ðeLN ;k −Sk Þ;

ð7Þ

so the labor biases can be considered either in combination or separately. Further, using the combined labor share equation SL = ( pLT LT + pLNLN)/pYY, the labor price or wage expression pL = ( pLT LT + pLNLN)/(LT + LN) = ( pLT LT + pLNLN)/L becomes, in share form, pL = (SLT + SLN)•pYY/L = pYY/L (SLTY + SLNY ). In turn, by the chain rule, the proportional impact on overall wages of a change in Xk, ∂ln pL/∂ln Xk, is: epL;Xk ¼ AlnpL =Aln Xk ¼ pY =pL L&ðASLT =Aln Xk &Y þ AY =Aln Xk &SLT þ ASLN =Aln Xk &Y þ AY =Aln Xk &SLN Þ ¼ pY Y =pL L&ðASLT =Aln Xk þ ASLN =Aln Xk þ Sk ðSLT þ SLN ÞÞ ¼ ðBLT;k þ BLN ;k þ Sk ðSLT þ SLP ÞÞ=SL :

ð8Þ

Again, therefore, the impacts on the individual labor components are implicit; e.g., εpLT,Xk = ∂ln pLT /∂ln Xk = (BLT,k + Sk SLT)/SLT. Note also that, for the TL functional form, the biases Bj,k are simply parameters such as BLTk = δLTk. Also, from (1), the estimated shares are simply the output elasticities such as SLT = εY,LT. Further, from (7), the input-specific impact of a change in Xk on, for example, LT, is εLT,k = ∂ln MPLT /∂ln Xk = (δLTk/SLT) + Sk. Thus, both relative (share) and absolute (marginal product or wage) impacts on LT, of (say) additional machinery or computers, Xk, can be computed to evaluate their complementarity or substitutability using estimates of the δLTk parameters and the εY,LT and εY,Xk elasticities. To compute these measures, we estimated our model (Eqs. (5) and (6)) by full information likelihood (FIML) system techniques to recognize the dependence of the error terms and the cross-equation constraints across equations. Although this generates estimates of the parameters and their standard errors, the output elasticities are combinations of these parameters and the data. Computing standard errors for these measures requires either evaluating the elasticity expressions at the averages of the variables in the data (the delta method), or using bootstrapping methods. Initially we calculated the output elasticities with respect to each input for each observation from the estimated parameters of our system of the estimating equations, and averaged them across the sample for each industry and year. For comparison purposes and to calculate standard errors we then evaluated the elasticities at the average values of the arguments of the production function using the delta method,11 and found discrepancies between the average values of the elasticities and the elasticities evaluated at the averages of the data.

11 Delta method is generalization of Central Limit Theorem, which is derived by using the Taylor series approximations. It is useful when one is interested in some function of a random variable rather than the random variable itself (See Gallant and Holly, 1980). We implement the delta method using TSP, which utilizes the parameter estimates from our system of model and their corresponding variance covariance matrix to evaluate the elasticities at average values of the inputs.

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

9

This could imply that the assumed normal approximation does not account for skewness of the marginal distributions of the parameters, in addition to the wide output variations across the sample for different plants (see Chapple et al., 2004). We thus tried using nonparametric bootstrap methods that do not require assumptions about the distributions and true values of parameters, to reduce potential statistical bias and provide more reliable standard error estimates. We used the bootstrap algorithm in TSP to randomly resample from our data, rerun the model with the resulting replications of the parameters, record the calculated elasticities from the bth resamples, and calculate the mean values and standard errors associated with the estimates across the whole sample and separately for each industry. More formally, we used nonparametric bootstrap methods introduced by Efron (1979) that obtain the sampling distribution of the statistic of interest by recycling the information in the sample.12 Assume that V has a cumulative distribution function G(v) = (V ≤ v) and that we have a data sample of size n from G(v). For a sample v = (V1,V2,…Vn) and an estimate of θˆ (V1, V2,… …,Vn) = θˆn, our goal is to select B resamples, where B is the total number of replications (each of size n), and to calculate the standard errors and the bias associated with the estimates. To accomplish this we draw a sample of size n from G, with replacement, to obtain the bootstrapped sample (V1⁎,V2⁎ ,… …,Vn⁎) (b = 1,2,… and the θˆ ⁎b1=2 from each of the samples.  …B) PB hˆ ⁎ ¯ˆ ⁎ i2 We then approximate the standard error as sˆeB ¼ − h =ðB−1Þ , where θˆ ⁎b is the statistic h b b¼1 P ⁎ B ⁎ ¯ ˆ b = B is the mean of the resampled values. calculated from the bth resample and hˆ ¼ b¼1 h¯ The accuracy of the estimators depends on the number of total replications (B). Efron and Tibshirani (1993) found that B = 250 produces a satisfactory approximation of the standard error, whereas for the bias B may need to be higher. We chose B = 500, which provides highly accurate approximations that more closely tracked the elasticity averages than those computed with the delta method. 4. Empirical results 4.1. The model estimates and sensitivity tests The parameter and standard error estimates for our final model are presented in Table 3, with the statistically significant estimates identified by asterisks; three asterisks implies statistical significance at the 1 percent level, two at the 5 percent level, and one at the 10 percent level. This final “preferred” model does not include the third order terms capturing the Dupuy and de Grip (2006) hypothesis of impacts on labor-capital relationships of computer use, or cross-effects between the imported capital or female labor share capturing the effects of, say, importing rather than developing technology on input substitutability and mix. It does, however, include all other input cross-terms, and thus is fully flexible in terms of inputs rather than imposing input separability. This model was supported by tests of the statistical significance of both the third and second order parameters, to determine whether a more restrictive model than Eq. (5) might justifiably represent production processes for our data.13 We carried out six tests. First, we tested the joint 12 Another method to draw samples from a given sample is jackknife, which was introduced by Quenouille (1956). The difference between jackknife and bootstrap is that the jackknife deletes a number of observations at each cycle of computation while bootstrap does random sampling of the sample observations at each cycle of computation (see Wu, 1986). 13 The possibilities that separability restrictions (resulting in a Cobb Douglas model if all restrictions are imposed), or a value added specification (that avoided potential materials endogeneity issues), might justifiably represent production processes were raised by anonymous referees. The Dupuy and de Grip (2006) hypothesis was also suggested as a testable hypothesis by an anonymous referee.

10

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

Table 3 Production function estimates (standard errors in parentheses) Independent Variables

Estimates

Independent Variables

Estimates

βM

0.044⁎ (0.025) 0.238⁎⁎⁎ (0.026) 0.182⁎⁎⁎ (0.022) 0.014 (0.024) 0.202⁎⁎⁎ (0.076) 0.040 (0.032) 0.242⁎⁎⁎ (0.032) 0.164⁎⁎⁎ (0.004) 0.020 (0.006) 0.027⁎⁎⁎ (0.004) 0.015⁎⁎⁎ (0.004) 0.040⁎⁎⁎ (0.008) 0.056⁎⁎⁎ (0.008) −0.043⁎⁎⁎ (0.004) −0.027⁎⁎⁎ (0.003) −0.003 (0.004) −0.014 (0.012) −0.010⁎⁎ (0.005) −0.056⁎⁎⁎ (0.005) 0.009⁎⁎ (0.004) 0.006 (0.004) 0.005 (0.013) −0.010⁎ (0.006) 0.011⁎ (0.006) −0.011⁎⁎⁎ (0.003)

γKMKC

−0.002 (0.011) 0.010⁎⁎ (0.005) −0.011⁎⁎ (0.005) −0.009 (0.013) −0.018⁎⁎⁎ (0.006) 0.013⁎⁎ (0.006) 0.034⁎⁎ (0.017) −0.043⁎⁎⁎ (0.017) −0.003 (0.007) 0.030 (0.068) 0.028 (0.042) 0.077⁎⁎⁎ (0.022) 0.199⁎⁎⁎ (0.028) 0.098⁎⁎⁎ (0.028) 0.034⁎ (0.019) 0.068 (0.053) −0.001 (0.009) 0.077⁎⁎ (0.058) 0.161⁎⁎⁎ (0.051) 0.167⁎⁎⁎ (0.049) −0.160⁎⁎⁎ (0.038) −0.092⁎⁎⁎ (0.034) −0.370⁎⁎⁎ (0.027) 2.856⁎⁎⁎ (0.104)

βE βKM βKS βKC βLT βLN δMM δEE δKMKM δKSKS δLTLT δLNLN δME δMKM δMKS γMKC δMLT δMLN δEKM δEKS γEKC δELT δELN δKMKS

δKMLT δKMLN γKSKC δKSLT δKSLN γKCLT γKCLN δLTLN βIS βFS βMedium βLarge β1995 β1996 βCentral Anatolia βBlacksea βMediterranean βAgean βMarmara βMeat Packing βTextiles βApparel α0

Notes: (1) ⁎⁎⁎Significant at the 1% level. ⁎⁎Significant at the 5 percent level. ⁎Significant at the 10 percent level.

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

11

significance of the third order terms, δjkC, and found that they were insignificant, so they were omitted. This indicates that the Dupuy and de Grip (2006) hypothesis is not supported by our data. Second, we tested for the significance of cross-terms between the IS and FS R components and the inputs, γrj, and again omitted them due to their insignificance. This suggests that the prevalence of imported technology and female labor in a developing country does not directly impact capital-labor relationships.14 Third, we tested for the joint significance of the input crossterms, δjk, and found that the restrictions of these terms to zero (and thus the hypothesis of a CobbDouglas model) were soundly rejected. This confirms the importance of recognizing input interactions, or the effects of plants' input composition on input “shares” or output elasticities, for our data. We then tested less restrictive models to determine whether only certain sets of crossterms among capital and the labor inputs might be important. For example, we tested whether the amount of non-technical labor, LN, does not affect capital machinery (KM) or skilled labor (LN) shares, so δNKM = δNT = 0. These tests were rejected for all combinations of the two labor inputs and capital except this particular example. The P-value for this test was just above 0.10 (0.11), providing weak support for the hypothesis (due to the very insignificant estimate of cross term between LT and LN, δNT) that did not seem sufficient to omit these cross-terms. We also tested for separability among combinations of labor and capital inputs, which is slightly less restrictive (it requires, e.g., terms such as δNKM and δNT to be the same but not necessarily equal to zero, weighted by their output shares, if non technical labor is separable from a combination of technical labor and machinery capital),15 with analogous results. Finally, we tried a value added production function specification, which is essentially a test of the separability of materials from other inputs.16 Although this specification is not directly testable since it is not nested, we found the magnitudes of the parameter estimates to be robust to this specification change, although their significance was reduced. This suggests that fixed proportions between materials use and output, which is implied by such a specification, is not supported by our data. The parameter estimates retained in the model and reported in Table 3 are generally significant, including most of the fixed effects and the input cross-terms. Exceptions are the estimated shift effects of the imported capital and female employment shares; even with the insignificant crossterms for these variables set to zero, the productive contributions of IS and FS are statistically insignificant. The explanatory power of the estimated model is also quite high for a primarily cross-section data sample; the R2s are 0.939 for the production function, and 0.345 for the share equation.17 14 All the cross-terms for IS and FS were very insignificant for the full specification. A test of joint insignificance for the IS parameters clearly supported the omission of these terms. A test of the joint significance of the FS terms was not as definitive; when other terms were dropped the cross-term between LT and FS became significantly negative. That is, plants that had a higher female share had a somewhat lower contribution of technical labor. 15 More specifically, as in Berndt and Christensen (1973), this requires that STγKMN−SKMγTN = 0, or αT/αKM = γTN/γKMN = γTi/γKMi (for all other inputs i, and where Sj = ∂ln Y/∂ln Xj as above). 16 As pointed out by an anonymous referee, evaluating the robustness of the results to this alternative specification could also be thought of as evaluating the possibility of the endogeneity of materials, so the robustness of the results is also supportive of our specification in this context. 17 It is worth noting that if the labor share equation was not estimated with the production function it made virtually no difference to the estimates, so the additional information gained from representing economic optimization, given the flexible nature of the production function, appears negligible. We retained this equation in the model, however, because it is conceptually useful and did not distort the results.

12

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

Table 4 Bootstrapped output elasticities and returns to scale, by industry (standard errors in parentheses)

Technical Labor, εY,LT Non-technical Labor, εY,LN Computers, εY,KC Machines, εY,KM Structures, εY,KS Materials, εY,M Energy, εY,E Returns to Scale, εY,X

Overall

Meat Packing SIC 3111

Textiles SIC 3212

Apparel SIC 3222

MotorVehicles SIC 3843

0.061⁎⁎⁎ (0.027) 0.109⁎⁎⁎ (0.018) 0.039⁎⁎ (0.017) 0.050⁎⁎⁎ (0.007) 0.010 (0.007) 0.723⁎⁎⁎ (0.010) 0.048⁎⁎⁎ (0.009) 1.003⁎⁎⁎ (0.018)

0.036⁎ (0.019) 0.069⁎⁎⁎ (0.022) 0.069⁎⁎ (0.028) 0.047⁎⁎⁎ (0.009) 0.010 (0.010) 0.787⁎⁎⁎ (0.021) 0.047⁎⁎⁎ (0.013) 0.996⁎⁎⁎ (0.019)

0.061⁎⁎⁎ (0.013) 0.109⁎⁎⁎ (0.018) 0.129⁎⁎⁎ (0.048) 0.050⁎⁎⁎ (0.007) 0.010 (0.007) 0.723⁎⁎⁎ (0.010) 0.048⁎⁎⁎ (0.009) 1.001⁎⁎⁎ (0.018)

0.068⁎⁎⁎ (0.014) 0.127⁎⁎⁎ (0.017) 0.038⁎ (0.021) 0.071⁎⁎⁎ (0.008) 0.003 (0.008) 0.689⁎⁎⁎ (0.012) 0.071⁎⁎⁎ (0.009) 1.029⁎⁎⁎ (0.010)

0.089⁎⁎⁎ (0.034) 0.154⁎⁎⁎ (0.051) 0.061⁎⁎⁎ (0.020) 0.070⁎⁎⁎ (0.022) 0.013 (0.017) 0.626⁎⁎⁎ (0.027) 0.057⁎⁎ (0.026) 1.009⁎⁎⁎ (0.044)

The parameter estimates for the fixed effects in D suggest that larger plants are more productive than smaller plants, that productivity is increasing over time, and that the motor vehicle industry is the most productive, followed by the textile, meat, and apparel industries, in terms of overall output per unit of input. The only R vector component for which cross-terms were significant and therefore was retained in the analysis is KC; computer use appears to have a direct significant productive impact (βKC is significantly positive), and its contribution is positively related to the use of technical labor and negatively related to non-technical labor (which we will return to below). 4.2. The first order elasticity estimates To explore the input contributions and their determinants, first consider the bootstrapped output elasticity (input productive contribution) and returns to scale estimates presented in Table 4 for the qualitative computer use variable, KC, as well as the levels of other capital, labor, and materials inputs, for the whole sample. Note that the only input that does not exhibit a statistically significant productive contribution is capital structures; although the εY,KS elasticity is positive and plausible at about one percent, it is only statistically significant in 1996. By far the largest share is for materials, at about 72 percent. Non-technical labor is next at about 11 percent, and then technical labor at about 6 percent. Computer use, machinery, and energy each exhibit productive contributions of about 5 percent. Note also that, allowing for fixed effects for medium and large plants, constant returns to scale appears to prevail for our sample of plants.18 The overall productive contribution of computer use implied by its elasticity, 0.05 or five percent, is close to the mean of that for the OECD countries (OECD, 2003). The contribution of machinery capital is also about five percent, consistent with its value share from Table 2, which is somewhat low compared to developed countries. In turn, the contribution of technical relative to When these fixed effects were omitted, statistically significant returns to scale of about three percent − εY,X = 1.03, was evident, so the shift variables appear to capture virtually all of the scale effects. 18

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

13

non-technical labor is greater than the share of technical as compared to non-technical labor evident from Table 2, indicating that skilled labor is a consequential productivity driver (its marginal product is greater than its cost). This is particularly important for a country like Turkey, which has less than 20 percent of the GDP per capita of the U.S. and about 25 percent that of the EU and OECD countries,19 much of which has been attributed to low labor productivity (Cotis, 2006). The combination of these estimated productivity effects suggest that increased capital intensity and use of computer technology combined with enhanced labor skills may help Turkey to break out of the “low starting point, low rates of catch-up” category of productivity relative to developed countries (OECD, 2003).20 The fact that this is true even for relatively low-skill industries such as those in by our data, which are representative of the industry mix in developing countries, underscores the importance of this finding. Variations in productive contributions of the inputs by industry are apparent from the industryspecific estimates presented in Table 4. The productive contributions of both technical and nontechnical labor are highest in the motor vehicles industry, followed in order by the apparel, textiles, and meat packing industries. The pattern is reversed for materials. The contribution of capital machinery is larger in the motor vehicles and the apparel as compared to the meat and the textiles industries (about 7 percent compared to about 5 percent), and the impact of computers is greatest in the textiles and lowest in the apparel industry (for which the energy share is the largest). Note also that the only industry that exhibits statistically significant (but small in magnitude) economies of scale, above the size fixed effects, is the apparel industry. 4.3. The second order bias estimates Next, turn to the second order effects that reflect the relationships of our primary interest – capital-skill interactions. The main measures reflecting these relationships – the bias measures – are captured directly by the cross-input parameter estimates from Table 3. Recall that these measures reflect the extent to which the input-specific effects of a change in Xk on Xj, εj,Xk, exceed the overall productive effects, εY,Xk. For example, consider BLT,KC = γKCLT, which represents, alternatively (given symmetry by Young's theorem), the impact on technical labor's productive contribution of computer use, or the impact on the productive contribution of computer use of more skilled labor. The significantly positive estimate of this bias implies complementarity between computers and technical labor, which may be thought of as technology-skill complementarity. The positive and significant δKMLT estimate similarly implies complementarity between capital machinery and technical labor. However, the significantly negative δKSLT and δKMKS estimates indicate substitutability between capital structures and technical labor in combination with capital machinery (and similarly for 19 That is, GDP per capita is less than 20 percent that of the U.S., although it is about 25 percent of the average of the OECD or EU countries (GDP per capital in the EU and OECD are very comparable, at about 73 percent of the U.S. level in the EU and 75 percent in the OECD countries as of 2002) (OECD, 2003). 20 Note that for the particular industries in this study, when we compared value-added per worker to the average for the EU countries using UNIDO data (United Nations Industrial Development Organization Industrial Statistics Database Indstat4, CD-Rom, 2005) at the 4-digit industry level, we found a much smaller (and statistically insignificant) difference although the wage difference was very large and significant. This suggests that part of the deviation between countries’ productivity is the industry mix, but also implies that our evaluation of the impacts on wages in these industries provides provocative information about how, even maintaining industry mix, productivity and wages may be enhanced by capitalskill synergies.

14

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

Table 5 Bootstrapped input contributions to labor wages (standard errors in parentheses) Computers, εpL,KC Machines, εpL,KM Structures, εpL,KS Materials, εpL,M Energy, εpL,E

Technical Labor ( pLT)

Non-technical Labor ( pLN)

10.880⁎⁎ (4.910) 4.398⁎⁎ (1.935) −2.050 (1.551) 2.409 (3.286) −0.568 (1.821)

− 4.517 (5.264) 1.043 (1.991) 2.298 (1.705) − 9.864⁎⁎ (5.010) 2.143 (2.170)

computer inputs although γKMKC is insignificant). The relationships between non-technical labor and the capital components, by contrast, suggest that LN is complementary with structural capital but substitutable with machinery and computer capital. These results support the notion that expanding capital intensity and particularly computerization could stimulate growth in Turkish manufacturing, contributing to “catch-up” with more developed countries. Although our estimates do not necessarily imply causation – that expanding computer capital and skilled labor in Turkey will result in greater productivity growth – they do show that plants that employ computer technology are more productive, and imply synergies between this technology and skilled labor as well as other capital machinery that could enhance Turkish productivity. Distinguishing just the wage effects from the overall share (quantity plus price) impacts implied by the bias measures requires using the estimated shares to compute the εpL,Xk elasticities defined in Eq. (8), which are presented for both types of labor in Table 5. These results confirm a (significantly) positive impact of computer use and capital machinery levels on the value and thus wages of technical labor ( pLT), although the estimated impacts on non-technical labor wages ( pLN) are insignificantly different from zero, suggesting that quantity rather than wage effects dominate share changes for non-technical labor. The only significant input composition effect on the wages for non-technical labor appears to be for materials; increasing materials use has a depressing effect on pLN, which might be expected since it suggests more purchases rather than production of intermediate inputs. These patterns again imply the potential for the expansion of capital – especially computer technology – and skilled labor components of production in Turkish manufacturing to have a consequential contribution to convergence of productivity and welfare toward the levels of developed countries. 5. Concluding remarks In this study we have used a production function model, augmented by a labor share equation, to measure the productive contributions of technical and non-technical labor and their dependence on capital levels and composition, for manufacturing plants in a developing economy. Overall, we find that higher capital machinery intensity and computer use are associated with a (significantly) greater skilled labor share and wage, which implies capital-skill or technology-skill complementary. The reverse is true for unskilled labor. By contrast, the shares of imported

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

15

capital and female employment have a negligible productive effect, either directly or through interactions with other inputs. These results suggest that synergies among skilled labor, capital intensity and computer use have the potential to help a developing country such as Turkey increase productivity and wages toward those in developed countries. This appears to be more important than the predominance of imported technology and female labor in this developing country. These patterns, which are generally consistent with those found for developed countries, prevail even given the differential input and industry composition of developing and developed countries. That is, these synergies are clearly apparent in the industries analyzed in this study, which are representative of industries dominating production in developing countries, even though production tends to be less skill- and capital-intensive in such a country than in the developed countries. It also appears that greater capital and skill intensity, although it has a depressing effect on unskilled labor share, does not significantly affect the wages of non-technical workers. Our results thus suggest that programs to increase capitalization and worker skills would likely contribute to both the growth and competitiveness of manufacturing industries and the welfare of manufacturing workers in a country such as Turkey. Acknowledgements We would like to thank Omer Gebizlioglu, Ilhami Mintemur and Emine Kocberber at the State Institute of Statistics in Turkey for allowing us access to the data for this study, and Erol Taymaz for helpful discussions about the data. We also are indebted to a co-editor and two anonymous referees for their helpful suggestions. Appendix A. Data appendix We use panel data for plants with more than 25 employees in the textile (manufacture of textile goods except wearing apparel, ISIC 3212), apparel (manufacture of wearing apparel except fur and leather, ISIC 3222), motor vehicle and parts (ISIC 3843), and meat processing (ISIC 3111) industries for 1995–1997. The data were collected by the State of Statistics in Turkey from Annual Surveys of Manufacturing Industries (ASMI), and classified based on the International Standard Industrial Classification (ISIC Rev.2). The SIS has conducted the Census of Industry and Business Establishments (CIBE) seven times — in 1927, 1950, 1963, 1970, 1980, 1985, and 1992, and carries out the ASMI for all establishments with 10 or more employees on an annual basis. Further, in every non-census year, addresses of newly opened and closed private establishments with 10 or more employees are obtained from the Chamber of Industry. From these data, the value of aggregate output (Y ) is defined as the value of total shipments, plus changes in inventories of finished goods and works-in-process. The value of non-energy materials (M ) is measured as expenditures on materials inputs, adjusted for changes in input stock levels. The value of energy (E) is computed by adding the value of fuel to that of electricity. The value of labor (L) is the total wage bill. There are two ways to compute capital stocks from the data. The first method uses gross investment, while the second uses the reported “depreciation fund” as a capital proxy. The first method is preferable because different companies might employ very different methods to compute depreciation, and any definition of depreciation at the plant level is unlikely to

16

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

correspond to the economic definition of capital. However, with data for only three years (the only years for which computer use data is available), it is difficult to construct a benchmark to compute the capital stock from the investment series. We thus use the depreciation allowances as a proxy for the capital input.21 To represent the quantities of these inputs from their value measures, for the non-labor inputs we divide their nominal values by corresponding price deflators to compute their constant dollar quantities at 1987 prices. For the labor input we have direct measures of the numbers of administrative, technical and production workers, and the share of female workers, For output we use a four-digit price deflator for each industry constructed by the SIS, based on the prices of the basket of commodities from each industry. We also use SIS deflators for the energy, capital and material inputs. The SIS calculates these deflators based on the Laspeyres formula, consistent with methods used by the Organisation for Economic Cooperation and Development (OECD).22 In addition to these input variables, we use year, industry, region and size indicators as control variables. Year dummies are included to capture macroeconomic shocks and the changes in the institutional environment. Four digit ISIC industry indicators are included to capture industryspecific shocks. Six regional dummies (for the East and Southeast Anatolian, Central Anatolian, Black sea, Agean, Marmara, and Mediterranean regions)23 are included to correct for exogenous disparities in productivity across regions, agglomeration effects that might affect productivity (Krugman, 1991; Porter, 1998), and development disparities due to different regional capabilities such as infrastructure, regulations, public services, and trade intensity. Finally, we use three size dummies (representing small plants with less than 50 employees, medium plants with between 50 and 100 employees, and large plants with more than 100 employees) to capture differences in production technology across different size plants. References Berman, E., Bound, J., Griliches, Z., 1994. Changes in the Demand for Skilled within U.S. Manufacturing Industries: Evidence from the Annual Survey of Manufacturing. Quarterly Journal of Economics 109 (2), 367–397. Berman, E., Bound, J., Machin, S., 1998. Implications of Skill Biased Technical Change: International Evidence. Quarterly Journal of Economics 113 (4), 1245–1279. Bernard, A.B., Jensen, J.B., 1995. Exporters, Jobs and Wages in U.S. Manufacturing Industries: 1976–1987. Brookings Papers on Economic Activity 67–119. Berndt, E.R., Christensen, L.R., 1973. The Translog Function and the Substitution of Equipment, Structures, and Labor in U.S. Manufacturing 1929–68. Journal of Econometrics 1, 81–114. Betts, J.R., 1997. The Skill Bias of Technological Change in Canadian Manufacturing Industries. Review of Economics and Statistics 79 (1), 146–150. Bresnahan, T., Brynjolfsson, E., Hitt, L., 2002. Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm Level Evidence. Quarterly Journal of Economics 117 (1), 339–376. Caroli, E., Van Reenen, J., 2001. Skill Biased Organization Change? Evidence from a Panel of British and French Establishments. Quarterly Journal of Economics 116 (4), 1449–1492.

21

Taymaz and Saatci (1997) use both methods but concluded that they produce very similar results. This deflation procedure may be problematic if there is excessive price and inflation rate variation across plants/ products within the industry. Griliches and Mairesse (1995) and Hall and Mairesse (1995) have noted that such procedures rely on the assumption that the “law of one price” for all firms holds — that is, all plants in the industry have the same output price and that their movements over time are similar, implying price-taking behavior at the firm level. The only way to correct for this requires detailed plant-level price data, which is typically available. However, the effect of violations to the “law of one price” is mitigated by utilizing appropriate weights for the product prices used in constructing the industry-specific deflators (which was done by SIS in constructing the deflator). 23 We combined the East and Southeast Anatolian region because there were not enough observations for East Anatolia. 22

M. Yasar, C.J. Morrison Paul / Labour Economics 15 (2008) 1–17

17

Chapple, W., Paul, C.J.M., Harris, R., 2004. Manufacturing and corporate environmental responsibility: cost implications of voluntary waste minimisation. Structural Change and Economic Dynamics 16 (3), 347–373. Cotis, Jean-Philippe, 2006. Economic Growth and Productivity. Presentation by OECD Chief Economist, the Government Economic Service Conference, Nottingham, 13–14 July. http://www.oecd.org/dataoecd/49/34/37179645.pdf. Doms, M., Dunne, T., Troske, K.R., 1997. Workers, Wages and Technology. Quarterly Journal of Economics 112 (1), 253–290. Dupuy, A., 2003. Will the Source of Demand Shifts Please Stand Up? Steady Demand or Acceleration? Working paper, Research Centre for Education and the Labour Market, Faculty of Economics and Business Administration, Maastricht University. Dupuy, A., de Grip, Andries, 2006. Elasticity of Substitution and Productivity, Capital, and Skill Intensity Differences across Firms. Economics Letters 90, 340–347. Efron, B., 1979. Bootstrap. Another look at the Jackknife. Annals of Statistics 7 (1), 1–26. Efron, B., Tibshirani, R., 1993. An Introduction to Bootstrap. Chapman and Hall, New York. Gallant, A.R., Holly, A., 1980. Statistical Inference in an Implicit, Nonlinear, Simultaneous Equation Model in the Context of Maximum Likelihood Estimation. Econometrica 48, 697–720. Goldin, C., Katz, L.F., 1998. The Origins of Technology-Skill Complementarity. The Quarterly Journal of Economics 114 (3), 693–732. Griliches, Z., 1969. Capital-Skill Complementarity. Review of Economics and Statistics 51 (4), 251–266. Griliches, Z., Mairesse, J., 1995. Production Functions: The Search for Identification. NBER Working Paper, No. 5067. NBER, Cambridge, MA. Hall, B.H., Mairesse, J., 1995. Exploring the relationship between R&D and productivity in French manufacturing firms. Journal of Econometrics 65 (1), 263–294. Haskel, J., Heden, Y., 1999. Computers and the Demand for Skilled Labour: Industry and Establishment-Level Panel Evidence for the U.K. Economic Journal 109, 68–79. Hellerstein, J.K., Neumark, D., 2004. Production Function and Wage Equation Estimation with Heterogeneous Labor: Evidence from a New Matched Employer-Employee Data Set. National Bureau of Economic Research Working Paper #10325. Hellerstein, J.K., Neumark, D., Troske, K.R., 1999. Wages, Productivity, and Worker Characteristics: Evidence from Plant-Level Production Functions and Wage Equations. Journal of Labor Economics 17, 409–446. Hollanders, H., ter Weel, B., 2002. Technology, Knowledge Spillovers and Changes in Employment Structure: Evidence from six OECD Countries. Labour Economics 9, 579–599. Krueger, A.B., 1993. How Computers Have Changed the Wage Structure: Evidence from Microdata, 1984–1989. Quarterly Journal of Economics 108 (1), 33–61. Krugman, P., 1991. Increasing Returns and Economic Geography. Journal of Political Economy 99, 483–499. Leiponen, A., 2005. Skills and Innovation. International Journal of Industrial Organization 23, 303–323. Machin, S., Van Reenen, J., 1998. Technology and Changes in Skill Structure: Evidence from Seven OECD Countires. Quarterly Journal of Economics 113, 1215–1244. OECD, 2003. STI Scoreboard 2003. http://www1.oecd.org/publications/e-book/92-2003-04-1-7294/. Porter, M.E., 1998. Clusters and New Economies of Competition. Harvard Business Review 77–90 (Nov–Dec). Quenouille, M.H., 1956. Notes on Bias in Estimation. Biometrika 43, 353–360. Revenga, A.L., 1992. The Impact of Import Competition on Employment and Wages in U.S. Manufacturing. Quarterly Journal of Economics 107 (1), 255–284. Siegel, D., 1997. The Impact of Computers on Manufacturing Productivity Growth: A Multiple-Indicators, MultipleCauses Approach. Review of Economics and Statistics 79 (1), 68–78. Taymaz, E., Saatci, G., 1997. Technical Change and Efficiency in Turkish Manufacturing Industries. Journal of Productivity Analysis 8, 461–475. Van Reenen, J., 1996. The Creation and Capture of Rents: Wages and Innovation in a Panel of U.K. Companies. Quarterly Journal of Economics 111 (1), 195–226. Wu, C.F.J., 1986. Jacknife, Bootstrap and Other Re-sampling Methods in Regression Analysis. Annals of Statistics 14, 1261–1295.