International Linkages and Productivity at the Plant Level:
Foreign Direct Investment, Exports, Imports and Licensing
Mahmut Yasar ∗
Department of Economics, Emory University, 306C Rich Building
Atlanta, Georgia 30322
Catherine J. Morrison Paul
Department of Agricultural and Resource Economics and the Giannini Foundation
University of California, Davis
One Shields Avenue,
Davis, California 95616
Abstract
Productivity growth may be affected particularly for developing countries by
international linkages or technology transfer. We evaluate relationships between
productivity and FDI, exports, imports and licensing for Turkish manufacturing plants in
the apparel, textiles, and motor vehicles industries. We assess performance premia
associated with these international technology transfer channels that control for plant size
and location. We then use a structural model to allow for plant-specific input composition
and interactions, estimated alternatively by quantile regression and semi-parametric
techniques to recognize plant heterogeneity and to accommodate simultaneity and
selection issues. Overall, we find that productivity is most closely related to foreign
ownership, especially for larger plants and in combination with other forms of technology
transfer, followed by exporting and then licensing.
Keywords: Importing; Foreign Direct Investment; Exporting; Firm Performance;
Technology Transfer
JEL Classifications: F10; F14; D21; L60
∗
Corresponding Author:
Tel: +1 404 712 8253; fax: +1 404 727 4639; E-mail: [email protected] (M. Yasar).
1. Introduction
Productivity growth determines the ability of an economy to improve its standard of living, and
is often considered to be the main source of cross-country income differences (Hall and Jones,
1999). An important issue in our increasingly global economic environment is thus whether
international linkages can enhance firms’ productivity and competitiveness. It may be
particularly important for developing or low/middle income economies such as Turkey to
identify how and to what extent productivity is related to international linkages that could narrow
the income gap from more developed countries.
Endogenous growth theory views innovation as the main source of productivity growth
(Romer, 1990, Lucas, 1988), although it may be associated with either internal or external
factors. In particular, studies have shown that international linkages or technology transfer may
be closely related to productivity growth (Coe and Helpman, 1995, Eaton and Kortum, 1999,
Keller, 2002). Countries such as Turkey that are still on a development path may especially rely
on the technology and knowledge produced by more developed countries rather than direct
investment in research and development. 1
Four main channels of international linkages appear in the literature. Foreign ownership
or foreign direct investment (FDI) is often considered the strongest conduit for international
technology transfer (Blomström and Kokko, 1998, Aitken and Harrison, 1999, Carr et al., 2001).
Learning by exporting has perhaps received the greatest attention (Kraay, 1997, Clerides et al.,
1998, Castellani, 2001, Bigsten et al., 2002, Girma et al., 2003). The role of technology
embodied in intermediate material and capital imports has been recognized (Grossman and
1
R&D expenditures for Turkey are, for example, only about 0.5 percent of GDP, compared to 2.4 percent for the
OECD countries, 2.6 percent for the U.S., and 2.9 percent for Japan (World Bank, 2004, average for 1996-2001).
1
Helpman, 1991, Xu and Wang, 1999, Eaton and Kortum, 2001). Foreign licensing has also been
considered (Eaton and Kortum, 1996), although it may not have a significant productive effect if
the best technologies are not available by license (World Investment Report, 2000).
These channels may have both separate and synergistic productive effects, as well as
linkages with internal factors such as input mix or size. Blomström and Kokko (1998), for
example, show that FDI may enhance host country firms’ productivity through knowledge flows
from cumulative R&D efforts in the foreign country, and of skilled employees and management
techniques across countries. However, productive effects may concurrently arise from exports
that move the domestic firms along the learning curve and imports of production technology by
the multinational enterprises. Augmenting labor force skills in the host country is also essential
for successful international technology transfer because it determines absorptive capacity
(Nelson and Phelps, 1966). Higher shares of technical or management workers would thus be
expected to be associated with greater productivity improvements from international linkages.
There are two primary hypotheses about how firm productivity is related to international
linkages. The first suggests that more productive firms self-select into, say, export markets,
because their characteristics make them better able to deal with the costs and complexities of
international markets. The second is that knowledge effects stem from exposure of exporting
firms to cutting-edge technology and managerial skills from their international counterparts.
Most empirical studies of export-productivity relationships support the self selection
hypothesis (Bernard and Jensen, 1995, Clerides et al., 1998, Aw et al., 2000, Delgado et al.,
2002), although others find learning-by-exporting (Kraay, 1997, Castellani, 2001, Bigsten et al.,
2002, Girma et al., 2003, Van Biesebroeck, 2005). Both of these effects may also be evident;
2
firms that participate in international markets may be inherently more productive but also
improve their productivity through international linkages (Yasar and Paul, 2005).
The role of firm heterogeneity in explaining relationships between productivity and
international technology transfer has also been explored theoretically. Melitz (2004) and Roberts
and Tybout (1997) find that more productive producers in an industry become exporters, and
Bernard et al. (2003) show that the heterogeneous efficiency underlying such choices implies
correlations across firms’ size, productivity, and export participation. Helpman et al. (2004)
confirm that more productive firms enter foreign markets, and that the most productive ones
engage in foreign direct investment.
Our objective is to examine the relationships between Turkish manufacturing plant
productivity (in the apparel, textile and motor vehicle industries) and all four technology transfer
channels, allowing for heterogeneous international linkages, plant characteristics, and input mix.
We first estimate within-industry proportional differences in performance indicators (premia),
controlling for plant characteristics such as location and size, for plants with and without
international linkages. We then directly examine production relationships underlying input
use/mix and international technology transfer through production function regressions that allow
a more structural analysis of plants’ productive processes and performance.
The structural analysis permits us to represent the productive effects of international
linkages in more detail than typical estimation of a simple functional relationship at the industry
or country level. However, plant heterogeneity raises implementation issues. For example, the
average productive effect of technology transfer may not well reflect the effects on different
plants with significant variations in capital intensity or size. Endogeneity issues inherent in
3
production function estimation are also exacerbated when attempting to measure and interpret
plant level productive impacts of international linkages. We deal with these issues in three ways.
First, representing the production technology by a flexible (translog) functional form
explicitly captures differential productivity patterns for plants with heterogeneous input
composition, reliance on international technology transfer, and other plant-specific
characteristics that our data allow us to identify. Second, using quantile regression techniques
allows for plant heterogeneity by representing the technology transfer effects at different points
of the conditional output distribution. Third, employing semi-parametric estimation including
international linkages as state variables accommodates simultaneity and selection bias.
Our overall conclusions are robust across methods. Our premia analysis suggests that
plants with a foreign ownership share (FDI) are the most productive, followed by plants that
export, especially in combination with other forms of international technology transfer. Plants
with international linkages are also larger, pay more, invest more, and hire more administrative
and technical workers. These findings are supported by the greater productivity effects
associated with FDI and exporting than licensing and importing, stronger productivitytechnology transfer relationships for larger plants, and higher productivity of especially larger
plants with more skilled labor, found from production function estimation. In addition, our
results show that a flexible functional form captures important input cross-effects, quantile
regression reflects productive discrepancies by plant size, and controlling for simultaneity and
selection bias increases the estimated productivity effects of technology transfer.
4
2. Data
Our analysis is based on unbalanced panel data for Turkish manufacturing plants with more than
25 employees, in the apparel 2 and textile (A&T) and motor vehicle and parts (MV&P) industries
from 1990-1996, 3 collected by Turkey’s State Institute of Statistics for the Annual Survey of
Manufacturing Industries. These industries combined generate about 50 percent of Turkey’s
manufacturing exports. 4 After cleaning the data to remove observations that had clearly
erroneous values or missing data 7024 observations remained, representing 1556 plants.
For international linkages, our data includes information on whether the plant exported
any products in the current year (EXP) and the sales share of exports (EXPS), whether the plant
imported any machine and equipment (IMP) and the investment share of those assets (IMPS),
whether the plant has any foreign ownership (FDI) and the foreign share (FDIS), and whether the
plant purchased any international technology through licensing (LIC). We also define dummy
variables representing the intersection of firms in multiple distribution channels as, e.g.,
EXPIMP=EXP•IMP to identify plants that are both exporters and importers.
Descriptive statistics of these variables, indicating the percent of plants falling into each
category for the dummy variables and the average plant values for the shares, are presented in
2
This industry includes all wearing apparel except fur and leather.
3
The export data are only available from 1990-1996, which restricted the sample for estimation purposes. These
data are similar to those used by Levinsohn (1993).
4
The textile, wearing apparel, and leather industry accounts for 35 percent of total manufacturing employment,
nearly 23 percent of wages, 20 percent of the output produced in the total manufacturing industry, and
approximately 48 percent of Turkish manufactured exports. The motor vehicle and parts industry accounts for 5
percent of total manufacturing employment, nearly 6.6 percent of wages, ten percent of the output produced in the
total manufacturing industry, and approximately 5.2 percent of Turkish manufactured exports.
5
Table 1. Note, for example, that in the A&T industry only 1 percent of plants with foreign
ownership did not utilize some other technology transfer mechanism, but 29 percent of the
observations showed exporting activity without any imports or FDI. 5 For another 6.5 percent of
observations the plants were both exporters and importers. The shares data further indicate, for
example, that the average share of foreign firms in the A&T industry is about two percent, and
the export intensity of the industry is about 27 percent.
For plant production variables, output (Y) is the deflated value of aggregate production,
with changes in inventory stocks taken into account. The material input (M) is deflated
intermediate materials expenditures, allowing for changes in material inventory stocks. The
energy input (E) is the deflated value of electricity and fuel. We use price deflators (in 1987
prices), also contained in the data, to divide the nominal values of these netputs to obtain their
real or “constant dollar” quantities. Labor quantity (L) is directly measured as total hours worked
in production (average hours worked times the number of employees). The capital input (K) is
measured by cumulating data on gross investment levels, deflated by a capital price index and
adjusted for depreciation (the perpetual inventory method). That is, K t = K t −1 (1 − δ ) + I t −1 , where
K t is the capital stock in period t, δ is the depreciation rate, and I t −1 is capital investment
between time periods t and t-1. 6
< INSERT TABLE 1 HERE >
For other plant-specific characteristics, we have data on the labor shares of
technical/engineering (TECHS), administrative (ADMS) and female (FS) workers. As shown in
Table 1, TECHS and ADMS are comparable for the two industries at about 2.5-4 and 15-20
5
Nearly 50 percent of the plants exported at some point during the sample period.
6
The assumed service lives for the fixed assets are 40 years for structures, 15 years for transportation equipment,
and 15 years for other machinery. The initial capital benchmark is computed by dividing a three-year average of
investment by the depreciation rate (Harper et al., 1989).
6
percent on average, but FS is much larger for the A&T industry. We also have data on the
expenditure shares of advertisement in total costs (ADV), inputs subcontracted to supplier plants
(SI), and output subcontracted by the other plants (SO). Our performance indicators include
average wages per employee (WAGE), capital machines/equipment per employee (MAC), total
investment per employee (TINV), the amount of administrative labor (LADM), the amount of
technical/engineering labor (LTECH), and total employment (EMP). As shown in Table 1, the
percentage of plants with less than 50 employees is 51 and 45 percent, with 50-100 employees
22 and 18 percent, and with more than 100 employees 26 and 36 percent, for the MV&P and
A&T industries, respectively.
3. Plant Performance of Firms with International Linkages
Our first objective is to assess the relationships between international linkages and
productivity/performance for these plants by evaluating correlation patterns. In particular, we
wish to measure the proportional differences between performance characteristics (Pit) of plants
that engage in international technology transfer (EXP, IMP, FDI, or LIC, in any combination),
and those that do not engage in these activities. We thus estimate the equation
10
R −1
T −1
j =1
r =1
t =1
ln Pit = β 0 + ∑ β jTTRANS jit + ∑ β r Dr + ∑ β t Dt + β EMP EMPit + uit
(1)
separately for the A&T and MV&P industries and for the logs of various plant-by-time (i,t)
characteristics reflecting differential productivity, capital, and employment capabilities.
More specifically, our ln Pit indicators include, in addition to total factor and labor
productivity (TFP, LP), wages paid and employment (WAGE, EMP), the amounts of
administrative and technical labor (LADM, LTECH), and machines and total investment per
employee (MAC, TINV). These variables were chosen due to hypotheses in the literature that
7
firms with international linkages employ and pay more, and have higher capital intensity, than
domestic firms (Bernard and Jensen, 1995, Bernard and Wagner, 1997).
The technology transfer variables in TTRANSjit are the dummy variables EXP, FDI, IMP,
LIC, EXPFDI, EXPIMP, EXPLIC, FDIIMP, FDILIC, and IMPLIC that distinguish groups of
firms with different types and combinations of international linkages. Regional dummies, 7 Dr,
capture productivity differences from great development disparities in Turkey in terms of
infrastructure, rule of law, public service quality, and density. Year dummies, Dt, reflect
macroeconomic shocks and changes in the institutional environment. EMP, a measure of plant
size, represents differences in the production technologies of different size plants. 8
The parameter βj thus indicates the average differences in ln Pit (percentage premia in
terms of the performance characteristic Pit) between plants in each technology transfer group and
the plants that do not have any international linkages, conditional on region, year, and size. We
present these estimated coefficients in Tables 2a and 2b for the MV&P and A&T industries,
respectively, with the variables ordered by their effects on total factor productivity (TFP). These
coefficients can be used to calculate median differences between the groups by taking the antilog
of the estimated coefficient on a group dummy, subtracting 1, and multiplying by 100 to obtain a
percentage difference (Halvorsen and Palmquist, 1980). For example, the coefficient on EXP in
the MV&P ln TFP regression is 0.152. The median TFP of exporting-only plants in the MV&P
industry is thus computed as (e β EXP − 1) * 100 = (1.1642-1)*100 ≈ 16.4; it is higher than that of
ˆ
the base group by about 16.4 percent.
7
There are 7 main geographic regions in Turkey: East Anatolia, South-East Anatolia, Central Anatolia, Black sea,
Agean, Marmara, and Mediterranean (the first two were combined due to limited observations for East Anatolia).
8
This is omitted when the ln Pit measure is based on overall employment or is on a per employee basis
8
Overall, most of the coefficients in the table are significant and positive, indicating that
plants in both industries with international linkages have higher productivity levels, are larger,
invest more, pay more, and hire more administrative and technical workers. We tested whether
the coefficients are equal between pair-wise groups, using F-tests, and found that the groups are
significantly different from each other, with few exceptions that include some combinations for
TFP and TINV in both industries and for LTECH in the A&T industry. 9
< INSERT TABLE 2a HERE >
In particular, significant total factor productivity differences across groups of plants are
evident from these measures. In the MV&P industry the most productive groups are plants that
have a foreign share and export (EXPFDI), plants that have a foreign share and obtain
technology through licensing (FDILIC), and plants that have a foreign share and import
machines and equipment (FDIIMP); the median TFP differences between these three groups and
the base group are 64.4, 45.5, and 45.4 percent, respectively. The top three groups in the A&T
industry are plants that export and have a foreign share (EXPFDI), plants with a foreign share
only (FDI), and plants that have a foreign share and import (FDIIMP), with median TFP
differences of 57.9, 57.6, and 41.9 percent from the base group. 10
< INSERT TABLE 2b HERE >
Our results for both industries support the finding of Bernard et al. (2003) that exporting
firms perform better and are larger than non-exporting firms. However, firms with foreign
ownership are even more productive relative to the base group, consistent with Helpman et al.
9
Details about these tests are available from the authors upon request.
10
If the export variable identifies plants that exported anytime during the time period rather than in the year under
consideration (by observation) the relative relationships are maintained but these premia become 44.6, 44.3 and 37.0
for the MV&P industry and 60, 54.6, and 40.2 for the A&T industry.
9
(2004), especially in combination with other forms of technology transfer. 11 That is, both
exporting and foreign-owned plants are more productive if they either import machines and
equipment or obtain technology through licensing, as shown by the coefficients on FDIIMP,
FDILIC, and EXPIMP, EXPLIC.
Helpman et al. (2004) also find that FDI sales relative to exports are larger in industries
with greater dispersion in firm domestic sales, which may arise from greater dispersion in
productivity. To assess the relative heterogeneity of our two industries we estimated their TFP
distributions by kernel density estimates, as shown in Figure 1. These distributions show broader
across-plant differences, or higher within-industry heterogeneity, for the A&T than the MV&P
industry. Combined with our finding that plants with a foreign share have higher median
productivity than comparison groups, and that these premia are greater in the A&T industry, this
also supports Helpman et al. (2004).
< INSERT FIGURE 1 HERE >
4. Econometric Analysis of Production Relationships
Our second objective is to estimate a translog production function for the plants in our data,
alternatively by ordinary least squares (OLS), quantile regression, and semi-parametric
regression techniques. This imposes more structure on our examination of the link between
productivity improvements and technology transfer, allowing the data to reveal more detailed
production structure relationships and their reliance on international linkages than simple
correlations. However, the great heterogeneity of the plants in our data raises questions about
inconsistent estimates across plants of widely differing sizes. Further, estimation of a production
function raises endogeneity/simultaneity problems, particularly if the results are interpreted in
terms of causal relationships.
11
This relationship is reversed, however, for labor productivity for the A&T industry.
10
Although endogeneity issues arise for any production function estimation, because firms
choose both output and inputs, when evaluating the link between productivity and international
spillovers the results must be interpreted with particular care in terms of correlation versus
causation. Both self selection (more productive plants decide to enter international markets) and
learning (international markets enhance plants’ productivity through knowledge transmission)
likely underlie observed production relationships (Yasar and Paul, 2005). 12
We accommodate such concerns to some extent by our flexible functional form, which
allows the data to reveal differential production patterns for heterogeneous plants, and our rich
dataset, which allows us to take into account many typically unobserved production factors. That
is, because the first derivatives of a flexible production function vary by observation the
estimates accommodate differential plant-level input mix and other characteristics. In addition,
our alternative econometric methods permit us to evaluate the robustness of our results to
specifications that control for size heterogeneity and simultaneity/selection issues.
More specifically, we assume that the production function Y = f(X,R,t) for Turkish
manufacturing plants can be represented by a translog approximation to the general function:
ln Yit = α0 + Σj βj ln Xjit + βt t + Σm βm Rmit + Σj γjt ln Xjit t + ΣmΣj γmj Rmit ln Xjit
+ .5 (ΣjΣk δjk ln Xjit ln Xkit + δtt t2) + φit ,
12
(2)
Yasar and Paul (2005) examined the causal relationships for productivity using matching methods and found that
exports, imports and foreign direct investment all cause differences in both labor and total factor productivity. Plants
with international linkages were more productive and larger before matching, implying self selection, but comparing
matched plants shows greater productivity for plants with the same characteristics except these linkages.
11
where i and t are plant and time subscripts (hereafter suppressed in most cases for notational
simplicity), Xj, ( j = K,L,E,M) is the jth input in the production process, and φit is a stochastic
error term. The international linkage variables included in R are the shares of foreign ownership,
exports and imports, and the licensing dummy (FDIS, EXPS, IMPS, and LIC). Internal R vector
components include technical, administrative, and female worker shares (TECHS, ADMS, FS);
subcontracted input and output shares (SI, SO); and advertising expenditures (ADV). Dummy
variables included in R designate plant size (small, medium, large), 13 year, region and industry.
With great plant heterogeneity, however, standard econometric estimates of (2) may not
be representative of the entire conditional output distribution (Mata and Machado, 1996), even
for a flexible functional form that captures variability through cross-effects. Heterogeneity may
cause the dependent variable and error term to be independently but not identically distributed,
and thus OLS estimates to be inefficient. Further, if the distribution of the dependent variable is
highly skewed, extreme observations will have significant impacts on the estimates. For our data
we do find great heterogeneity in plant size, and that the distribution deviates from normality. 14
That is, quantile regression methods correct for inefficiencies in OLS estimation resulting
from this type of plant heterogeneity. They are relatively robust to departures from normality
because they place less weight on outliers in the distribution of the dependent variable. For our
application they allow us to represent differential productivity-international linkage relationships
13
The size dummies, where plants with less than 50 employees are deemed “small” and those with 100 or more
employees “large,” capture some scale-related technology differences.
14
More specifically, we used tests outlined by D’Agostino et al. (1990) to evaluate the distribution of Y and ln Y,
and found both to be positively skewed and leptokurtic (skewness = 0.000 and kurtosis = 0.000). We also applied a
Jarque-Bera test to an OLS model to examine the normality of the conditional distribution of residuals, and the
hypothesis of normality was rejected at the 0.01 significance level.
12
at different points of the conditional output distribution, or across very different size plants, as
suggested by the theoretical models of, e.g., Bernard et al. (2003). 15
Quantile regression can be expressed in the general form (Koenker and Basset, 1978,
Buchinsky, 1998) ln Yit = z 'it βθ + φit with Qθ (ln Yit / zit ) = z 'it β θ , where for our purposes z is a
vector of all the explanatory variables in (2) and β the vector of associated parameters. Qθ
denotes the θ th conditional quantile of ln Yit given z it and the size of operations. Estimates for
any quantile of the distribution of ln Y conditional on z are obtained by changing θ continuously
from zero to one, and using linear programming methods to minimize the sum of weighted
absolute deviations. This provides plant-size-specific information about the effects of regressors
on the dependent variable, rather than the effects of regressors at the conditional mean of the
dependent variable, as with OLS. The same measures may thus be computed as for OLS, but
they will differ by quantile – for different size plants. 16
Estimation of (2) also raises questions about simultaneity and selection biases. Since
plants choose input, technology transfer mechanisms and output simultaneously, unobserved
plant characteristics may cause the error term for (2) to be correlated with the arguments of the
production function. This violates the orthogonality of the error term with the inputs and
technology transfer variables, so standard OLS techniques are biased and inconsistent.
15
See Dimelis and Louri (2002) and Mata and Machado (1996).
16
Equality of the estimates for plants in the various quantiles can be tested using the variance-covariance matrix of
the system of quantile regressions, estimated by bootstrapping techniques. See Bassett and Koenker (1982) and
Hendricks and Koenker (1991) for further information on the computation of the test statistic. For a review of
quantile regressions see Koenker and Hallock (2001) and Buchinsky (1998).
13
If one could include all omitted variables in the regression the error term would be
orthogonal to the inputs and technology transfer variables and the coefficients would
appropriately reflect plant productivity relationships. Although our data do capture key
production characteristics, leaving less unobserved heterogeneity than in many studies, not all
production determinants are observable. Also, lagging the technology transfer variables may help
to mitigate the problem, but this is not a complete solution since unobserved factors may affect
both production in year t and the technology transfer variables in year t-1.
We can also use the semi-parametric model of Olley and Pakes (1996) 17 to control for
remaining simultaneity and selection bias. 18 Applying this method requires assuming that a plant
chooses its variable inputs conditional on beginning of the period state variables, including a
productivity shock (assumed to follow a first-order Markov process) and the existing capital
stock. For our application FDIS, EXPS and IMPS also become state variables. Expected
productivity and resulting decisions about whether to produce or exit the industry and to invest in
capital are thus functions of these state variables.
That is, plants that experience a positive productivity shock in period t-1 will remain in
the marketplace and increase their capital investment in period t, so the inputs and technology
variables are correlated with the productivity shock, resulting in simultaneity bias. The first step
of the Olley and Pakes approach accommodates this bias by expressing the productivity shock as
a second-order polynomial function of the investment and state variables (FDIS, EXPS, IMPS
17
Attempts were also made to accommodate endogeneity by dynamic GMM procedures, but the lagged values of
the international linkage variables were weak instruments. The Sargan test for overidentifying restrictions suggests
that the endogenous variables dated t-2 and earlier are not valid instruments.
18
A similar procedure is implemented by Van Biesebroeck (2005) to examine the learning by exporting effects for
Sub-Saharan African manufacturing firms.
14
and K). Further, a plant with higher levels of state variables will expect higher future profitability
at current productivity levels and thus have less incentive to exit, resulting in selection bias. This
is dealt with in a second step by estimating the probability of a plant staying in the market as a
second order polynomial series in lagged investment and state variables. In the last step a semiparametric series estimator is approximated by a second-order polynomial to obtain consistent
coefficients on the state variables. The resulting estimates thus control for unobserved permanent
differences across plants, sample selection, and simultaneity.
5. Econometric Results
In preliminary estimation of equation (2) we included a full set of interaction terms between all
input and plant-characteristic variables, but ultimately omitted many from the model because
they were individually and jointly insignificant. In particular, although interaction terms were
significant for inputs, they were generally insignificant for the R vector components, so the γmj
terms were set to zero. The output implications from differences in these variables may thus be
assessed simply through their estimated first order coefficients, βm.
In addition, we initially alternatively used qualitative (dummy) and quantitative (share)
measures for the international linkage variables. The qualitative variables show whether or not
the plant engaged in international technology transfer (FDI, EXP, IMP, LIC), and the
quantitative measures reflect the intensity of technology transfer activity (EXPS, IMPS, FDIS).
For both specifications the international linkage (as well as employment characteristic) variables
were lagged by one year. We report the results using the quantitative measures because our
primary results were similar, the intensity measures better represent the extent of international
linkages, and the qualitative variables cannot be used as state variables for the OP model.
15
However, we comment below on some informative differences between these results and those
from a specification based on qualitative technology transfer variables.
< INSERT TABLE 3 HERE >
Table 3 presents the coefficient estimates of equation (2) for the trade and labor variables
we are focusing on, pooled across the industries for our alternative stochastic specifications, with
one, two, and three asterisks indicating statistical significance at the 1, 5, and 10 percent
significance levels. 19 The first column presents the estimates for the OLS regression, the second
to sixth columns the estimates for quantile regressions (at the 0.10, 0.25, 0.50, 0.75, and 0.90
output level quantiles), and the last column the estimates from the Olley-Pakes (OP) model.
The coefficient estimates for IMPS, EXPS and FDIS reflect productivity differences for
plants with a greater share of technology transfer. The OLS coefficient estimates on the
technology transfer variables have the expected positive signs and are statistically significant,
with the FDIS and export effects generally larger than those for licensing and imports, and the
FDIS effect greater than that for EXPS, consistent with the evidence from our performance
premia regressions and the literature overall. More specifically, the coefficient for FDI implies
that plants with a 0.1 (10 percent) larger foreign ownership share in year t-1 are about 2.2 percent
more productive in year t, conditional on the control variables. Similarly, the estimates indicate
19
The remaining production function estimates are available from the authors upon request. We initially tested
whether pooling was justifiable by doing separate estimation for the MV&P and A&T industries to see whether
significant differences were apparent, and found that they were not. We then pooled the industries but tested to see
whether coefficients for the two industries varied. The test results, F (4, 5237) = 0.14 and Prob > F = 0.9673, shows
that the regression coefficients do not significantly differ by industry. We interpret these results as evidence that, as
found by Bernard et al. (2003), within-industry heterogeneity dominates between-industry heterogeneity.
16
that a 10 percent higher export or imported technology share is associated with about 1.5 or 0.6
percent greater productivity, respectively.
In addition, plants that licensed international technology have an estimated (significantly)
higher output level of nearly 12 percent relative to those that did not. This effect is not directly
comparable to the other estimates, however, because it is based on a dummy variable and thus
reflects the implications of the existence rather than an increasing share of technology transfer.
For estimation alternatively based on qualitative measures of all technology transfer variables the
estimated productivity effects of the other international linkages were also larger. They indicated
that plants with a foreign share or that were exporters were about 13.5 percent, and those that
imported technology about 5 percent, more productive than those that did not. The estimate for
licensing was somewhat smaller (but still significant) in that specification, implying about 6.5
percent greater productivity for firms that licensed technology.
In turn, the quantile regression estimates generally span the OLS estimates and thus
maintain the relative rankings of the technology transfer effects, 20 but the extent of their
variability indicates the importance of recognizing the plant size differences. The coefficient on
FDIS is significant across the entire conditional output distribution but varies from about 0.10 to
0.26 from the lower to the higher quantiles, with a slightly lower estimate for the very largest
plants than for the 25th quantile, relative to 0.22 for OLS. For exporters the range is 0.13 to
0.165, with the peak at the 50th quantile, compared to 0.15 for OLS. Both plants with greater
import shares and those that obtain technology through licensing exhibit a relatively high
associated productivity for the plants in the 25th quantile, although the greatest licensing effect
20
This would be expected since they represent an additional dimension of variation. Although this is not true for
import share, such patterns have been found by others such as Mata and Machado (1996).
17
appears for the largest plants. 21 The statistical significance of size-related variations in
productivity effects was supported by hypothesis tests of differences in parameter estimates
between pairs of quantiles and across all quantiles. 22
The quantile regression results broadly suggest a greater productivity effect of
international linkages for larger plants, which is most definitive for FDI and least for IMP.
Alternative estimates based on qualitative international technology transfer variables revealed,
however, a stronger monotonic tendency for all channels, consistent with Bernard et al. (2003)
and Helpman et al. (2004). These relative differences suggest that although larger domestic-only
plants are less productive relative to comparable plants with international linkages, their share of
international activity need not be as large as for smaller plants to gain from these linkages.
Further, when simultaneity and selection biases are accommodated by OP estimation we
would expect the coefficients on the variable inputs to be lower, but on the state (technology
transfer and capital) variables to be higher, than with OLS. The estimates presented in Table 3
are consistent with these a priori expectations, although the relative effects of international
linkages are not changed; the productivity effect of FDI is the greatest, followed by exporting,
licensing, and importing. Specifically, the coefficients on the foreign share, export share, and
21
We also estimated a model including both in-plant and industry-level shares of export and foreign sales and
import investment, to reflect the possibility of spillovers from technology transfer. The magnitudes and significance
of the plant-level variables were essentially maintained with this adaptation, and only the export spillover variable
exhibited any statistical significance, particularly for the larger quartiles. The FDI variable was also highly
significant for the OP model, but not the quantile regressions (except the 0.9 quantile) and OLS.
22
The results of this test are available from the authors upon request.
18
imported investment share are 0.224, 0.151 and 0.059 for OLS and 0.318, 0.169 and 0.084 for
the OP model. 23
The more internal but related productivity effects of labor composition may also be
evaluated from the estimates in Table 3. For example, productivity may be related to a more
skilled workforce because adoption of new technology requires skilled labor. 24 This is broadly
supported by our estimated coefficients on technical and administrative personnel shares, which
are positive and the most significant for the OP model and the highest quantiles. In reverse, since
developing countries often utilize female labor for less technical jobs, female labor share may
indicate low skill levels. Our coefficient estimates confirm a negative productivity relationship
for female share, especially for the larger plants.
6. Concluding Remarks
In this paper we examined the relationships between international linkages through four
technology transfer channels and plant productivity in the Turkish apparel, textile, and motor
vehicles manufacturing industries. Our evaluation was based on both regressions of performance
indicators on international linkage variables, and estimation of a production function that
captures more production structure and interrelationships.
Our reliance on plant-level data, for our production function analysis in particular, raises
heterogeneity issues. Standard productivity measurement techniques implicitly assume that
plants within an industry share common productivity relationships, whereas they may actually
23
The biases were also as expected for the variable input and capital variables.
24
For example, Tan and Batra (1995) show that joint ventures with foreign companies can facilitate the transfer of
technology because they are implemented by foreign management and accompanied by training. Hanson and
Harrison (1999) note that plants that obtain new technology through licensing agreements and imported materials
need to hire workers with high skill levels in order to fully utilize the technology and imports.
19
differ in important ways. This potential heterogeneity is emphasized by Bernard et al. (2003),
Melitz (2004), and Helpman et al. (2004), who theoretically examine the effect of within-sector
heterogeneity of firms engaging in international activities.
Our use of a translog functional form generalizes standard Cobb-Douglas production
function models to capture plant heterogeneity through output-input interactions underlying
productivity and scale effects. Incorporating data on key plant characteristics normally
unobserved for econometric analysis also reduces difficulties arising from heterogeneity.
However, at least two issues remain that were controlled for using alternative stochastic models.
First, great variation in plant size could veil differences in productivity relationships for
different size plants, which we deal with using quantile regression techniques. Second,
endogeneity or simultaneity problems arise when including technology transfer variables in the
production function, which we accommodate using an Olley-Pakes (1996) semi-parametric
estimation procedure. The primary results do not differ substantively by stochastic specification.
However, as expected from theory, estimates of the productivity relationships with international
linkages vary by quantile, with stronger ties often apparent for the larger quantiles, and are
underestimated compared to the OP estimates.
Our results for all our estimation models and methods confirm that firms with
international linkages are more productive. In particular, they support findings in the recent
literature that foreign ownership (FDI) and exporting are positively related to plant-level
productivity, and that the FDI effect is greater than that for exports. Further, licensing and
importing technology are significantly related to productivity, and to the productivity
implications of FDI and exporting. In addition, internal plant characteristics such as the share of
skilled labor enhance the productive role of international linkages.
20
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 Jonathan Eaton and two
anonymous referees for their constructive and useful suggestions.
21
References
Aitken, B., Harrison, A., 1999. Do domestic firms benefit from foreign direct investment?
Evidence from Venezuela. American Economic Review 89, 605-618.
Aw, B., Chung, S., Roberts, M., 2000. Productivity and turnover patterns in the export market:
Firm level evidence from Taiwan and South Korea. World Bank Economic Review, 14, 65-90.
Bassett, G., Koenker, R., 1982. An empirical quantile function for linear models with iid Errors.
Journal of the American Statistical Association 77, 407-415.
Bernard, A.B., Eaton, J., Jensen, J. B., Kortum, S., 2003. Plants and productivity in international
trade. American Economic Review 93, 1268-1290.
Bernard, A. B., Wagner, J., 1997. Exports and success in German manufacturing.
Weltwirtschaftliches Archive 133, 134-157.
Bernard, A.B. Jensen, J. B., 1995. Exporters, jobs and wages in US manufacturing: 1976-1987.
Brookings Papers on Economic Activity, Microeconomics 1995, 67-119.
Bigsten, A., Collier, P., Dercon, S., Fafchamps, M., Gauthier, B., Gunning, J. W., Habarurema, J.,
Oduro, A., Oostendorp, R., Pattillo, C., Soderbom, M., Teal, F., Zeufack, A., 2002. Do African
manufacturing firms learn from exporting?. Oxford University, Centre for the Study of African
Economies Working Paper Series, WPS/2002-09.
Blomström, M., Kokko, A., 1998. Multinational corporations and spillovers. Journal of Economic
Surveys 12, 247-277.
Buchinsky, M., 1998. Recent advances in quantile regression models: A practical guide for
empirical research. Journal of Human Resources 33(1), 88-126.
Carr, D. L., Markusen, J. R., Maskus, K.E., 2001. Estimating the knowledge-capital model of the
multinational enterprise. American Economic Review 91(3), 693-708.
22
Castellani, D., 2001, Export behavior and productivity growth: Evidence from Italian
manufacturing firms, Mimeo, ISE-Università di Urbino.
Clerides, S., Lach, S., Tybout, J., 1998. Is learning-by-exporting important? Micro dynamic
evidence from Colombia, Mexico, and Morocco. Quarterly Journal of Economics 53, 903–947.
Coe, D., Helpman, E. 1995. International R&D spillovers. European Economic Review 39,859-887.
D’Agostino, R.B., Balanger, A., D’Agostino Jr., R.B., 1990. A suggestion for using powerful and
informative tests of normality. The American Statistician 44(4), 316-321.
Delgado, M., Fariñas, J., Ruano, S., 2002. Firm productivity and export markets: A non-parametric
approach. Journal of International Economics, 57, 397-422.
Dimelis, S., Louri, H., 2002. Foreign ownership and production efficiency: A quantile regression
analysis. Oxford Economic Papers 54, 449-469.
Eaton, J., Kortum, S., 1996. Trade in ideas: Patenting and productivity in the OECD, Journal of
International Economics 40, 251-271.
Eaton, J., Kortum, S., 1999. International patenting and technology diffusion: Theory and
measurement, International Economic Review 40, 537-570.
Eaton, J., Kortum, S. 2001. Trade in capital goods, European Economic Review 45 (7), 1195-1235.
Girma, S., Greenaway, D., Kneller, R., 2003. Export market exit and performance dynamics: A
causality analysis of matched firms, Economics Letters, 80, 181-187.
Grossman, G., Helpman, E., 1991. Innovation and growth in the global economy, MIT Press,
Cambridge, MA.
Hall, R., Jones, C., 1999. Why do some countries produce so much more output per worker than
others? Quarterly Journal of Economics 114, 83-116.
23
Halvorsen, R., Palmquist, R., 1980. The Interpretation of dummy variables in semilogarithmic
equations. American Economic Review, 70(3), 474-475.
Hanson, G. H., Harrison, A. 1999. Trade, technology and wage inequality in Mexico. Industrial
and Labor Review 52, 271-288.
Harper, M., Berndt, E., Wood, D., 1989. Rates of return and capital aggregation using alternative
rental prices, in: Jorgenson, D. W., Landau, R. (Eds.), Technology and Capital Formation,
Cambridge, MA: The MIT Press, pp. 331-372.
Helpman, E., Melitz M. J., Yeaple, S.R., 2004. Export versus FDI with heterogeneous firms.
American Economic Review 94, 300-316.
Hendricks, W., Koenker, R., 1991. Hierarchial spline models for conditional quantiles and the
demand for electricity. Journal of American Statistical Association 87, 58-68.
Kraay, A., 1997. Exports and economic performance: Evidence from a panel of Chinese
enterprises, Working Paper, World Bank, Washington, DC.
Keller, W., 2002. Geographic localization of international technology diffusion. American
Economic Review 92, 120-142.
Koenker, R., Basett, G., 1978. Regression Quantiles. Econometrics 46, 33-50.
Koenker, R., Hallock, K. F., 2001. “Quantile Regression. Journal of Economic Perspectives 15(4),
143-156.
Levinshon J., 1993. Testing the imports-as-market-discipline hypothesis. Journal of International
Economics 35, 1-22.
Lucas, R. E., 1988. On the mechanics of economic development planning. Journal of Monetary
Economics 22(1), 3-42.
24
Mata, J., Machado, J.A.F., 1996. Firm start-up size: A conditional quantile approach. European
Economic Review 40, 1305-1323.
Melitz M., 2004. The impact of trade on aggregate industry productivity and intra-industry
reallocation. Econometrica 71(6) 1695-1725.
Nelson, R., Phelps, E., 1966. Investment in humans, technological diffusion and economic growth.
American Economic Review 56, 69-75.
Olley, G. S., Pakes, A., 1996. The dynamics of productivity in the telecommunications equipment
industry. Econometrica 64(6), 1263-97.
Roberts, M., Tybout, J.R., 1997. The decision to export in Colombia: An empirical model of entry
with sunk costs. American Economic Review 87, 545-564.
Romer, P. M., 1990. Endogenous technological change. Journal of Political Economy 98, 71-102.
SIS. State Institute of Statistics. Ankara, Turkey: www.die.gov.tr.
Tan, H. W., Batra, G., 1995. Enterprise training in developing countries: Incidence, productivity
effects, and policy implications, Private Sector Development, World Bank, Washington, DC.
Van Biesebroeck, J., 2005. Exporting raises productivity in Sub-Saharan African manufacturing
firms. Journal of International Economics, 67(2), 373-391.
World Investment Report, 2000. Cross-border mergers and acquisitions and development, UN
Conference on Trade and Development, 1st edition.
World Bank. 2004. World Development Indicators. CD-ROM. Washington, D.C.
Xu, B., Wang, J., 1999. Capital goods trade and R&D spillovers in the OECD. Canadian Journal of
Economics 32, 1258-1274.
Yasar, M., Paul, C. J. M., 2005, International technology transfer and productivity: evidence from a
matched sample. Manuscript.
25
Table 1. Descriptive statistics, plant trade, labor and size characteristics
(percent of observations for dummy variables, plant averages for share and size variables)
Dummy variables
FDI (only)
EXP (only)
IMP (only)
LIC (only)
EXPFDI
EXPIMP
EXPLIC
FDIIMP
FDILIC
IMPLIC
NONE
Shares
FDI
EXP
IMP
Technical workers
Administrative workers
Female workers
Size
<50 employees
50-100 employees
>100 employees
MV&P Industry
(1511 observations)
0.79
12.51
7.04
2.63
0.59
5.46
3.95
1.45
3.03
8.03
54.51
A&T Industries
(5513 observations)
0.96
28.62
4.93
0.14
0.64
6.47
0.23
0.89
0.13
0.46
56.51
5.20
6.69
22.70
3.96
18.81
8.19
1.67
27.20
19.69
2.61
15.20
46.59
45.36
18.27
36.37
51.31
22.44
26.25
26
Table 2a: Percentage differences between groups, MV&P Industry
EXPFDI
FDILIC
FDIIMP
LIC
IMPLIC
EXPLIC
EXP
EXPIMP
IMP
FDI
EXPFDI
FDIIMP
FDILIC
LIC
IMPLIC
EXPLIC
EXP
EXPIMP
IMP
FDI
ln TFP
ln LP
ln WAGE
ln EMP
0.497*
(0.127)
0.375*
(0.061)
0.374*
(0.084)
0.322*
(0.062)
0.289*
(0.042)
0.287*
(0.054)
0.152*
(0.032)
0.118*
(0.045)
0.104*
(0.039)
0.095
(0.110)
0.993*
(0.257)
1.523*
(0.113)
1.601*
(0.166)
0.839*
(0.120)
1.382*
(0.073)
1.027*
(0.102)
0.362*
(0.059)
0.696*
(0.087)
0.437*
(0.075)
0.689***
(0.214)
0.662
(0.183)*
1.173
(0.081)*
1.052
(0.118)*
0.717
(0.085)*
1.297
(0.052)*
1.057
(0.073)*
0.348
(0.042)*
0.687
(0.064)*
0.305
(0.054)*
0.853
(0.153)*
1.420
(0.302)*
2.268
(0.133)*
2.217
(0.195)*
1.187
(0.140)*
2.710
(0.086)*
1.736
(0.120)*
0.916
(0.069)*
1.392
(0.103)*
0.746
(0.088)*
1.368
(0.252)*
ln LADM
ln LTECH
1.715
(0.374)*
2.486
(0.241)*
2.833
(0.165)*
1.650
(0.174)*
3.161
(0.106)*
2.297
(0.148)*
1.194
(0.085)*
1.653
(0.127)*
0.776
(0.109)*
1.770
(0.311)*
0.888
(0.322)*
1.385
(0.208)*
1.991
(0.142)*
1.013
(0.153)*
2.253
(0.093)*
1.454
(0.128)*
0.640
(0.076)*
1.162
(0.110)*
0.752
(0.097)*
0.988
(0.268)*
ln MAC
1.024
(0.484)**
1.989
(0.313)*
0.712
(0.214)*
0.853
(0.243)*
1.965
(0.139)*
0.678
(0.196)*
0.224
(0.121)***
1.968
(0.175)*
1.513
(0.145)*
1.804
(0.404)*
ln TINV
0.023
(0.574)
1.768
(0.305)*
0.519
(0.225)**
0.617
(0.280)**
1.722
(0.139)*
0.010
(0.214)
0.089
(0.136)
1.587
(0.169)*
1.403
(0.144)*
1.798
(0.444)*
27
Table 2b: Percentage Differences between Groups, A&T Industry
EXPFDI
FDI
FDIIMP
EXP
FDILIC
EXPIMP
IMPLIC
EXPLIC
LIC
IMP
ln TFP
ln LP
ln WAGE
ln EMP
0.457*
(0.092)
0.455*
(0.075)
0.350*
(0.078)
0.261*
(0.017)
0.234
(0.206)
0.222*
(0.031)
0.204***
(0.104)
0.201
(0.151)
0.136
(0.193)
0.046
(0.035)
0.950*
(0.169)
0.365**
(0.128)
0.289**
(0.145)
0.899*
(0.028)
1.074*
(0.367)
1.096*
(0.055)
1.132*
(0.201)
0.851*
(0.260)
0.385***
(0.222)
0.541*
(0.059)
0.424
(0.074)*
0.426
(0.056)*
0.513
(0.064)*
0.196
(0.012)*
0.412
(0.162)**
0.403
(0.025)*
0.853
(0.090)*
0.428
(0.114)*
0.156
(0.098)
0.178
(0.026)*
1.103
(0.142)*
0.533
(0.107)*
1.278
(0.121)*
0.574
(0.023)*
1.576
(0.308)*
1.229
(0.046)*
2.113
(0.168)*
0.930
(0.218)*
0.665
(0.186)*
0.842
(0.049)*
ln LADM
ln LTECH
ln MAC
ln TINV
1.267
0.366
0.400
0.456
(0.184)*
(0.185)**
(0.263)
(0.383)
0.830
0.036
0.951
0.811
FDI
(0.139)*
(0.158)
(0.205)*
(0.272)*
1.391
0.195
1.572
1.109
FDIIMP
(0.157)*
(0.148)
(0.217)*
(0.223)*
0.834
0.071
0.386
0.321
EXP
(0.030)*
(0.035)**
(0.047)*
(0.058)*
2.128
0.433
2.044
1.014
FDILIC
(0.399)*
(0.464)
(0.549)*
(0.642)
1.552
0.571
1.940
1.614
EXPIMP
(0.060)*
(0.060)*
(0.091)*
(0.092)*
2.682
1.011
1.713
1.235
IMPLIC
(0.218)*
(0.205)*
(0.306)*
(0.312)*
1.441
0.088
0.000
0.173
EXPLIC
(0.282)*
(0.269)
(0.431)
(0.524)
0.936
-0.169
1.326
1.220
LIC
(0.241)*
(0.253)
(0.517)**
(0.525)**
0.951
0.298
1.843
1.390
IMP
(0.064)*
(0.068)*
(0.092)*
(0.096)*
Notes: (1) Robust standard errors are in parentheses. *Significant at the 1% level. ** Significant at the 5% level.
***Significant at the 10% level. (2) The independent variables include year, size, and region dummies (the
regressions in which the dependent variable is employment or is on a per employee basis do not include the size
variable). Dependent variables are in natural logs. (3) The base group is the plants that do not have any international
linkages.
EXPFDI
28
Table 3: Output contributions (production function estimates, standard errors in parentheses)
Independent
Variables
OLS
Estimates
0.224*
(0.043)
0.151*
Export Share
(0.016)
0.059*
Import Share
(0.018)
Licensing
0.116*
Dummy
(0.031)
Engineering
0.145***
Share
(0.077)
Administrative 0.135**
Share
(0.057)
-0.078*
Female Share
(0.028)
Apparel
-0.294*
Industry
(0.022)
Motor
0.134*
Vehicles and
(0.025)
Parts Industry
0.230*
ln EMP
(0.038)
FDI Share
OP
Estimates
0.10
0.097***
(0.055)
0.128*
(0.021)
0.045***
(0.024)
0.096**
(0.041)
-0.149
(0.100)
0.118
(0.086)
0.003
(0.039)
-0.359*
(0.029)
Quantile Regression Estimates
0.25
0.50
0.75
0.204*
0.198*
0.255*
(0.042)
(0.039)
(0.051)
0.163*
0.165*
0.136*
(0.015)
(0.014)
(0.020)
0.055*
0.052*
0.053*
(0.018)
(0.017)
(0.023)
0.117*
0.094*
0.111*
(0.030)
(0.028)
(0.039)
0.081
0.099
0.239**
(0.073)
(0.069)
(0.093)**
0.098***
0.151*
0.166**
(0.058)
(0.051)
(0.067)
-0.001
-0.031
-0.115*
(0.028)
(0.026)
(0.035)
-0.351*
-0.342*
-0.272*
(0.021)
(0.020)
(0.026)
0.90
0.236*
(0.074)
0.106*
(0.029)
0.014
(0.034)
0.139**
(0.057)
0.328*
(0.116)
0.135***
(0.069)
-0.123**
(0.050)
-0.176*
(0.037)
0.200*
(0.034)
0.162*
(0.024)
0.117*
(0.023)
0.090*
(0.030)
0.081*** 0.143*
(0.044)
(0.028)
0.127**
(0.054)
0.162*
(0.038)
0.232*
(0.035)
0.222*
(0.048)
0.281*
(0.068)
0.318*
(0.065)
0.169*
(0.024)
0.084*
(0.024)
0.140*
(0.032)
0.457*
(0.118)
0.172*
(0.066)
-0.071**
(0.035)
-0.306*
(0.025)
0.211*
(0.044)
Note: *Significant at the 1 percent level. **Significant at the 5 percent level. ***Significant at the 10 percent level.
29
MV&P
A&T
1.2
.9
.6
.3
-3
-1
1
3
LNTFP
Figure 1. Kernel Density Estimate of the Distribution of Total Factor Productivity in the Apparel
and Textile and Motor Vehicle and Parts Industries
30