FOREIGN TECHNOLOGY TRANSFER AND PRODUCTIVITY:
EVIDENCE FROM A MATCHED SAMPLE
Mahmut Yasar and Catherine J. Morrison Paul*
ABSTRACT
This study examines the causal effects of alternative forms of foreign technology transfer
on the productivity of Turkish manufacturing plants through exporting, foreign direct investment
and importing. We use propensity score matching techniques that limit implicit assumptions
about plant homogeneity imbedded in standard estimates of such effects, and that control for
selection bias (i.e. the endogeneity of international involvement variables). We find positive
impacts of technology transfer through exporting, foreign direct investment, and importing on
both total factor and labor productivity of the plants. Foreign direct investment has the greatest
productivity impact, followed by exporting and importing, respectively. Our results also support
the hypothesis in the literature that internationally involved plants exhibit better performance
than domestic plants before matching.
Keywords: Foreign Technology Transfer, Foreign Direct Investment; Exports; Imports; Plant
Performance; Causality
JEL Classifications: F10; F14; D21; L60
* Department of Economics, Emory University, and Department of Agricultural and Resource
Economics, University of California, Davis, and Giannini Foundation.
Acknowledgements: The authors would like to thank Omer Gebizlioglu, Emine Kocberber, and
Ilhami Mintemur at State Institute of Statistics in Turkey for allowing us access to the data for
this study.
I. Introduction
Productivity deviations across plants, industries, and countries can stem from a variety of factors
in addition to differing rates of technical change. In particular, international diffusion of
technology, especially across countries at different stages of development, may affect
productivity. Such effects are consistent with the premise of the endogenous growth literature
that trade enhances growth because international transactions give firms access to new
technology. This facilitates learning and contributes to firms’ productivity and thus
competitiveness with other global players in world markets.
Channels through which technology diffusion may occur include importing intermediate
and capital goods that embody new technology, “learning by exporting” products that imitate
other countries’ technology, and foreign direct investment (FDI) that increases knowledge flows
from foreign R&D and training of domestic employees. Various studies have examined the
hypothesis that firms, industries or countries that transfer technology through these mechanisms
exhibit greater productivity. For example, Coe and Helpman (1995), Xu and Wang (1999), and
Eaton and Kortum (2001) evaluate the impact of importing; Kraay (1997), Castellani (2001),
Bigsten et. al. (2002), Wagner (2002), and Girma, Greenaway and Kneller (2003) consider the
effect of exporting; and Blomström and Kokko (1998), Aitken and Harrison (1999), and Carr,
Markusen, and Maskus (2001) address the influence of FDI or ownership. 1 Firm level studies in
this literature typically have examined the productivity effects of either exporting or FDI, or
both, and suggest that FDI is the most significant channel. Industry studies have evaluated all
three channels, as well as the licensing of technology (Eaton and Kortum, 1996), and tend to find
significant individual productivity effects of FDI, exporting, and importing, but not licensing.
1
Most of these studies use standard least squares techniques to estimate technology
transfer effects by including a (typically dummy) variable reflecting the existence of foreign
linkages (say, importing), in an equation representing the production technology and thus
productivity (output production for a given amount of inputs). The (constant) coefficient on the
technology transfer variable therefore represents the increase in output producible from a given
amount of inputs due to the internalization of foreign technology. However, because the general
question targeted in this literature is whether productivity differs for plants that do or do not
transfer technology, issues arise about how productivity is measured, and how best to represent
the counterfactual of a plant not transferring technology when it is observed to do so.
Specifically, estimates of production function parameters capturing the implicit weights
on inputs’ productivity contributions may suffer from simultaneity and selection biases. In
addition, these methods at least implicitly assume homogeneity of the technology transfer effect
across the distribution of explanatory variables, since they impose a linearity restriction on the
outcome equation, and assume rather than establish causation. The estimates may therefore
confound productivity differences from technology transfer with those arising from other plantspecific characteristics that may have caused it to choose to transfer technology.
In this study we measure productivity using a semi-parametric model to generate
unbiased estimates. We also directly compare the productivity of plants with similar
characteristics that have chosen to transfer or not to transfer technology, using a non-parametric
propensity score matching technique that does not impose a linearity restriction and thus allows
us more explicitly to represent the counterfactual of a plant not transferring technology to
determine causal inference. We apply these methods to examine the impacts of foreign
technology transfer channels on Turkish manufacturing plant productivity.
2
Various productivity measures could be used to evaluate technology transfer impacts.
One such measure is labor productivity, which can simply be computed as output per unit of
labor input. However, a more representative measure of resource use is total factor productivity,
which reflects output production per unit of aggregate inputs. Adding up the inputs used for
production requires weighting the contributions of each input, typically using output share
measures from estimation of a production function relationship. Such estimation may, however,
be subject to simultaneity and selection bias due to unmeasured plant-level heterogeneity, the
correlation of inputs with the realization of unobserved productivity shocks, and a relationship
between profitability and capital stock that affects the choice of a plant to exit the market. We
thus use the three-step semi-parametric method developed by Olley and Pakes (1996) to generate
unbiased productivity estimates.
In turn, propensity score matching methods collapse a vector of plant characteristics to a
single “score” representing the plant’s propensity to transfer technology (Rosenbaum and Rubin,
1983), based on a probit regression relating plants’ decisions and characteristics. The method
used ensures that plants matched to compare productivity effects have the same distribution of
characteristics independently of their choice to transfer technology (they satisfy the “balancing
hypothesis”), so the choice is statistically random and the estimates unbiased with respect to self
selection. It also compares only observations with propensity scores within the intersection of
scores achieved for both plants that transfer and do not transfer technology (they are restricted to
the “common support”).
Based on these scores, we alternatively match plants using “nearest neighbor” and
“kernel” techniques that impose different weighting schemes. The nearest neighbor technique
attaches a weight of one to the non-transferring plant with the closest propensity score to the
3
transferring plant (with replacement so the non-transferring plant can be used as a match more
than once). This matches each plant that does transfer technology with the one plant with the
closest propensity score, but may result in some poor matches. The kernel matching technique
instead matches each transferring plant with a weighted average of all non-transferring plants,
where close matches have a large and poor matches a small weight.
Such matching techniques have been applied extensively in labor economics to examine
“treatment effects” in the presence of selection bias, which occurs when the individual
characteristics of program participants differ from non-participants (Rosenbaum and Rubin,
1983, Rosenbaum, 1984, Heckman, 1997, Heckman, Ichimura, and Todd, 1997, 1998, Heckman
et al., 1998, Heckman, Lalonde and Smith, 2000, and Angrist, 2003). In the terminology of this
literature, computation of the effect of “treatment” (in our case technology transfer) requires
comparing the observed performance (productivity) of a “program participant” (plant that
transferred foreign technology) with the performance that would have resulted had that
participant not been treated (had the plant not transferred any technology).
This study contributes to the foreign technology transfer literature by using these methods
to provide empirical evidence of productivity effects caused by technology transfer for
manufacturing plants in a low to middle-income economy (Turkey) that might be expected to
benefit from the diffusion or transfer of technology. 2 More specifically, using this methodology
we can ask the following questions more directly than is possible using more standard
techniques: (i) does the higher productivity of plants that transfer foreign technology through
these channels result from this activity or from certain plant characteristics regardless of
technology transfer?; (ii) what performance might we expect from plants that do not transfer any
foreign technology, based on their given characteristics, if they did transfer foreign technology
4
through these channels?; and, (iii) does transferring foreign technology improve the productivity
of plants? We address these questions using two different productivity measures (labor and total
factor productivity) and two matching methods (nearest neighbor and kernel), for the three
primary foreign technology transfer channels raised in the literature – imports, exports, and FDI.
We find that plants’ export market participation, foreign ownership, and imports of
machines, equipment and materials, imply higher total factor and labor productivity. This
confirms the positive relationships often found in the literature, while establishing causal
inference by explicitly formulating the counterfactual as the productivity level the plant would
have had without any foreign technology transfer through these channels.
II. The Literature
Coe and Helpman (1995) was an early study in the literature on the productivity impacts of
foreign technology transfer, with a focus on trade spillovers through importing. Technology
transfer through this channel would be expected to affect a country’s productivity if importing
materials and capital from countries with high levels of accumulated technological knowledge
transfers this technology through R&D embodied in the inputs, and learning associated with their
use. In their analysis of OECD countries and Israel, Coe and Helpman (1995) find that both
domestic and foreign (through imports) R&D stocks augment total factor productivity, with
foreign technology dominating in the smaller, and domestic in the larger (G7), countries.
Similarly, Xu and Wang (1999) find significant machinery stock import spillovers on
productivity and growth for these countries, and Eaton and Kortum (1996) find that most of this
growth stems from technology imported from the U.S., Germany, and Japan.
In addition to these macro studies, Keller (2002a,b) finds significant imported technology
productivity impacts among the G7 countries using industry level data. The few studies that use
5
firm level data to evaluate this relationship generate somewhat conflicting results. Kraay,
Isoalaga, and Tybout (2001) find no evidence of productivity effects from importing, using firm
level data from Colombia, Mexico and Morocco, and Keller and Yeaple (2003) find weak
evidence of such effects for U.S. manufacturing firms.
By contrast, a number of studies have examined the relationship between exports and
productivity using firm- or plant-level data, including Bernard and Jensen (1995, 1999) for the
United States; Aw, Chung, and Roberts (2000) for Taiwan and Korea; Clerides, Leach, and
Tybout (1998) for Colombia, Mexico, and Morocco; Kraay (1999) for China; Bernard and
Wagner (1997, 2001) and Wagner (2002) for Germany; and Girma, Greenaway and Kneller
(2003, 2004) for the U.K. 3 The idea underlying this literature is that exporting enhances firm
productivity via a “learning-by-exporting effect.” That is, such firms learn about and adopt
international best practice production methods, receive feedback from international clients and
competitors that may improve product offerings, and benefit from other knowledge spillovers.
Although this argument is intuitively plausible, there is little empirical evidence
supporting it; only Kraay (1999), Castellani (2001), Bigsten et al. (2002), and Van Biesebroeck
(2003) find a significant export effect. Most empirical studies, such as Bernard and Jensen
(1999), Clerides, Lach, and Tybout (1998), Aw, Chung, and Roberts (2000) and Delgado,
Fariñas, and Ruano (2002), find that the dominant causation is in the opposite direction; more
productive firms self-select into export markets. This self-selection theory suggests that
exporting plants are more productive than non-exporting plants not because of technology
transfer benefits from export activity, but because they are more productive from the outset.
The bulk of the literature on the relationship between FDI and productivity is based on
industry level data. For example, Caves (1974), Globerman (1979), and Blomström (1986) use
6
data on Canadian and Australian, Canadian, and Mexican manufacturing industries, respectively,
to show that industries with higher foreign shares are more productive. Some studies have,
however, used firm or plant level data to examine such relationships, such as Blomström and
Wolff (1989) for Mexico, Haddad and Harrison (1994) for Morocco, Doms and Jensen (1998)
and Helpman, Melitz and Yeaple (2004) for the U.S., and Aitken and Harrison (1999) for
Venezuela and Indonesia. Productivity effects through this technology transfer channel are
expected to stem from flows of knowledge associated with cumulative foreign R&D efforts, and
of skilled employees and management techniques. Studies in this literature tend to find higher
productivity levels but not growth rates for firms with more foreign ownership, suggesting
productivity convergence between domestic and foreign-owned firms (Blomström and Wolff,
1989), especially for smaller firms (Aitken and Harrison, 1999). 4
A problem with this literature, however, as well as most studies on import and some on
export technology transfer effects, is that existing studies typically examine the association
between FDI and productivity rather than causation (Aitken and Harrison, 1999). This may
overstate the estimated productivity effects, as is suggested by the literature on exporting. In this
study, therefore, we wish not only to address the relationships between plant-level productivity
and all three channels of foreign technology transfer, but also to establish the causality
underlying this relationship, which we accomplish using propensity score methods.
III. The Empirical Approach
Treatment Effects
Let IMPit ∈ {0,1} be an indicator variable of whether plant i imported any machines, equipment
and material at time t, and let τ iI1 denote the value of the plant’s productivity (τ) for plants that
purchased foreign technology (I1 denotes IMP=1) and τ iI0 the same plant’s productivity level
7
had it not imported any technology (I0 denotes IMP=0). The productivity effect of importing for
plant i (the treatment effect) can then be defined as τ iI1 - τ iI0 . The fundamental problem of causal
inference for this effect is that τ iI0 is unobservable, because one firm cannot be an importer and a
non-importer at the same time (Holland, 1986).
We then define the effect of importing – the average effect of treatment (imports) on the
treated (importing plants), ATT – as in Holland (1986) and Heckman, Ichimura and Todd (1997):
ATT = E [τ iI1 - τ iI0 | IMPi = 1] = E [τ iI1 | IMPi = 1] - E [τ iI0 | IMPi = 1] .
(1)
Causal inference relies on the representation of the counterfactual (the last term in equation 1),
which is the productivity level the plant would have had if it did not import. This counterfactual
is often estimated using the productivity level of plants that did not import, E [τ iI0 | IMPi = 0]. 5
However, this does not permit direct causal inferences because variations in firm characteristics
likely affect observed differences in the productivity of importers and non-importers.
That is, if there is a deviation in productivity between importers and non-importers we
cannot infer that importing technology was its cause, because plants that imported could have
already been more productive prior to importing due to other unmeasured factors. This implies
selection bias; firms that transfer technology likely differ in some observable manner from those
that do not. Therefore, to improve causal inferences, the goal is to obtain a good estimate for the
unobserved component – the average outcome in the untreated state. This requires setting up an
experiment that essentially “randomizes” the treatment (technology transfer) of plants, such that
the “selection” of treated plants is uncorrelated with either observable or unobservable
characteristics, and comparisons across the units are unbiased, as elaborated below.
Similarly to the case of importing technology, the average productivity or treatment
effects on plants of foreign direct investment or exporting can be defined as:
8
ATT = E [τ iF1 - τ iF0 | FDI i = 1] = E [τ iF1 | FDI i = 1]- E [τ iF0 | FDI i = 1]
(2)
ATT = E [τ iE1 - τ iE0 | EXPi = 1] = E [τ iE1 | EXPi = 1]- E [τ iE0 | EXPi = 1]
(3)
where FDI it ∈ {0,1} and EXPit ∈ {0,1} are indicator variables for foreign plants and exporters,
respectively, τ iF1 and τ iE1 are the productivity levels of foreign and exporting plants, and τ iF0 and
τ iE0 are the productivities of domestic and non-exporting plants. The estimated counterfactuals in
this case would thus be E [τ iF0 | FDI i = 0] and E [τ iE0 | EXPi = 0] , and causation of productivity
deviations through these technology transfer channels can be inferred by matching foreign or
exporting plants with domestic or non-exporting plants.
In addition to addressing questions about the average treatment effects on the treated,
ATT, we can calculate the average treatment effects for the whole population of treated (say,
importing) plants as ATE = E[τ iI1 - τ iI0 ] , and for the controls (non-treated plants) as
ATU = E[τ iI1 - τ iI0 | IMPi = 0] . That is, ATE measures allow us to examine whether transferring
foreign technology has a positive impact on the overall or average productivity of plants – the
average outcome differences between the treated (importing) and control plants. ATU is by
contrast the potential impact of transferring technology on the plants that did not transfer any
technology but had the “propensity” to do so, or the average difference between the potential
productivity of the plants who did not import any technology if they had imported technology
( τ iI1 ), and the actual productivity outcome that they attained ( τ iI0 ).
In sum, our primary measures of interest are the average productivity or treatment effects
of technology transfer – the mean effects of importing or exporting, or being a foreign firm, as
compared to not importing or exporting, or being a domestic firm, for the plants that transferred
9
technology through these channels. The sample analogs of the ATT, ATE, and ATU measures
representing these effects are computed as: 6
ATT =
1
1
1 N D1 D0
D1
D0
(
);
(τ i − τ i ); and ATU =
τ
−
τ
ATE
=
∑
∑
i
i
N i =1
N0
N1 i|Di =1
∑(y
i|Di = 0
1
i
− y i0 ),
where D stands for IMP, FDI or EXP, and N1 = ∑iD i and N 0 = ∑i(1 − D i ) are the numbers of
treated (IMP=1, FDI=1, EXP=1) and control (IMP=0, FDI=0, EXP=0) units.
Computing such effects requires matching “treated” plants – plants that transferred
foreign technology through importing, exporting, or FDI – with untreated plants that have very
similar characteristics. That is, constructing the counterfactual involves selecting the appropriate
comparison group of non-participants, such that exposure to the treatment is random. Hence, if
there is a difference in productivity it can be attributed to firms’ technology transfer rather than
other characteristics, and if there is no difference it is reasonable to infer that plants that
transferred foreign technology would have been more productive even if they did not do so.
When experimental data is used this problem is handled by randomly assigning
individuals into a treatment group that participates in the program and a control group that does
not. This randomization ensures a complete balancing of all relevant observable and
unobservable characteristics across the treatment and control group; potential outcomes are
independent of treatment status (independence). If the treated and controls are identical except
for participation in the program, the difference in the outcome variable can be considered solely
attributable to the program (Lalonde, 1986).
However, with non-experimental data for which there is no randomized control group,
the goal is to find a control group that is as similar to the treated as possible, by using matching
estimators to approximate the counterfactual outcome (Blundell and Costa Dias, 2000). The use
10
of matching to construct a valid control group is based on the identifying assumption that,
conditional on all relevant observable covariates Z, the potential outcomes are independent of
treatment (conditional independence) – so treatment status is random. 7
Propensity Scores and Matching Methods
Since matching plants on the basis of n characteristics may be infeasible particularly if n is large,
methods have been proposed to summarize such characteristics into a scalar variable or
“propensity score.” 8 The propensity score is defined by Rosenbaum and Rubin (1983) as the
conditional probability of receiving treatment (transferring technology) given pre-treatment (no
technology transfer) characteristics:
pi(Zi) ≡ Pr{Di = 1/Zi} = E(Di/Zi} ,
(4)
where Di = {0,1} is the indicator of “exposure to treatment” (transfer of technology here), IMPi,
FDIi, or EXPi, and Zi is the vector of the ith plant’s characteristics on which the match is made.
For our purposes, the propensity score pi(Zi) is computed from a probit regression of a
binary variable indicating whether or not a plant transferred technology in time t through a
specific channel on relevant plant characteristics in time t-1. The plant characteristics we use as
explanatory variables (Z) for the plant’s decision to transfer technology include lagged values of
the logarithms of wages (ln W), logarithms of capital intensity (ln KI), logarithms of the number
of workers (ln LP), the share of imported machines and equipment in total investment (IMPS),
the share of foreign ownership (FDIS), labor shares of administrative (AWS) and technical (TWS)
workers, the export share (production of total output that is directed to foreign markets, EXS), the
subcontracted input share (input share of subcontracts to the supplier plants, SCI), and the
subcontracted output share (share of output subcontracted by other plants, SCO). We also include
year, and industry and regional dummy variables.
11
Linking (4) with (1), (2), or (3) (deriving ATT, ATE or ATU from pi(Zi)) requires that
plants satisfy the balancing property formalized in Becker and Ichino (2002), where the matching
of plants is “balanced” if observations with the same propensity score have the same distribution
of observable (and unobservable) characteristics independently of treatment status. This implies
that the decision to transfer is random; treated and “control” units are observationally identical
on average. The choice of probability model to estimate the propensity score E(D/Zi) = F(h(Zi)),
where F(·) is the cumulative distribution and h(Zi) a function of covariates with linear and higher
order terms, is based on the need to satisfy this property. That is, the choice of linear and higher
order terms for estimation must verify the balancing property that plants with the same
propensity score have the same distribution of the observed covariates. 9
The quality of the matches used to estimate ATT, ATE, and ATU is also improved by
restricting matching to plants that fall in the “common support,” defined as the observations
whose “propensity score belongs to the intersection of the supports of the propensity score of
treated and controls” (Becker and Ichino, 2002). For our application some plants that transfer
technology have no comparable plants, so estimation of technology transfer effects in the
absence of common support results in coefficients that reflect the effects of both treatment and
pre-treatment variables. We eliminate this bias by dropping treated plant observations with
propensity scores higher than the maximum or less than the minimum of the controls.
Even after verifying the balancing property, however, and including only the treated
plants that fall in the common support, it is unlikely to find plants that transfer and do not
transfer technology with exactly the same propensity score. To overcome this difficulty, we
alternatively use nearest neighbor and kernel matching techniques to match the plants.
12
With the nearest neighbor matching technique, for each importing (treated) plant one
non-importing (non-treated) plant with the closest propensity score is selected. That is, this
methods assigns a weight of one for the nearest comparison unit in terms of the propensity score,
and zero to all other observations. We implement this technique with replacement since a treated
plant can be a best match for more than one non-treated plant. The problem with this matching
technique is that some matches may be poor; for some treated plants the nearest neighbor
(matched control) may have a very different propensity score. However, this plant and its match
will still contribute equally to the estimation of the treatment effect with other plants that have a
much closer match, potentially biasing the overall results. 10
With kernel matching, by contrast, plants that transfer technology (say, importers) are
matched with a weighted average of all non-importers with weights that are inversely
proportional to the distance between the propensity scores of importers (treated) and nonimporters (controls). Formally, the weighting function is a (Gaussian) kernel density. A large
weight is thus given to close matches and a small one to poor matches.
The main advantage of using matching techniques rather than parametric estimation of
productivity as a function of treatment (technology transfer) is that they do not rely on functional
form or distributional assumptions in the estimation of the causal effects. 11 Furthermore,
matching recognizes the problem of common support – that for some treated there are no
comparable plants – and thus that the technology transfer effect estimated in the absence of
common support results in coefficient estimates that include both the effect of treatment and the
pre-treatment variables. Matching methods, that impose conditional independence through the
balancing property and common support, reduces these potential biases.
13
Productivity Measures
Estimating productivity differences between matched plants that do and do not transfer foreign
technology also requires unbiased measures of productivity. Labor productivity (LP) is typically
defined as output (Y) per unit of labor input (L), Y/L, or in log-form as ln LP = ln Y – ln L. This
measure can be computed directly from data on Y and L. Total factor productivity (TFP) is
similarly defined as output per unit of aggregate input (X), Y/X, or ln TFP = ln Y-ln X. However,
production function estimates required to sum the inputs into an aggregate input X based on
output elasticities or “shares” may suffer from simultaneity and self selection problems. That is,
ln TFP is often computed by approximating ln X from estimation of a production function such
as ln Y = Σj βj ln xj (for J inputs xj). Computing the productivity residual ln TFP thus involves
weighting the inputs xj by the estimated output elasticities βj, which may be biased due to
simultaneity and self selection. For our purposes, ln TFP was therefore estimated using a semiparametric approach that controls for such biases (Olley and Pakes, 1996).
This approach assumes that plants decide at the beginning of each period whether to
continue to participate in the market. If the plant exits, it receives a liquidation value of Φ. If it
does not, it chooses variable inputs and realizes profits, conditional on the beginning of the
period state variables – a productivity indicator or shock, Ωit , and the capital stock, K it . We
further assume that expected productivity is a function of current productivity and capital,
E[ Ωi,t +1 | Ωit , K it ] and that the firm’s profit is a function of Ωit and K it .
Plant i’s decision to maximize the expected discounted value of net future profits can
then be represented by the Bellman equation:
Vit ( K it ,Ωit ) = Max{Φ, SupΠ it ( K it ,Ωit ) − C ( I it ) + ρ E[Vit +1 ( K it +1 ,Ωit +1 ) / J it ]} ,
14
(5)
where Πit (⋅) is the profit function (current profits as a function of the state variables), C (⋅) is the
cost of current investment, ρ is the discount factor, and E[⋅ /J it ] is the plant’s expectations
operator conditional on information ( J it ) at time t. A plant thus exits the market if its liquidation
value (Φ ) exceeds its expected discounted returns. The solution to this equation generates a
Markov Perfect Equilibrium strategy defining rules for exit and investment choices.
Specifically, the plant will decide to stay in the market (χ t = 1) or exit the market
( χ t = 0) if its productivity is greater or less than some threshold, respectively, subject to its
current capital stock K it . This exit rule is formalized as:
⎧⎪1 if Ωit ≥ Ω it ( K it )
−
χt = ⎨
,
⎪⎩0 Otherwise
(6)
where the state variable Ωit is assumed to follow a first-order Markov process.
The plant’s decision to invest in additional capital, I it , also depends on K it , and Ωit :
I it = I ( Ωit , K it ).
(7)
This investment decision equation implies that future productivity is increasing in the current
productivity shock, so plants that experience a large positive productivity shock in period t will
invest more in period t+1.
Based on these exit and investment decision rules, we can specify a production function
to estimate total factor productivity in an unbiased manner. We assume that the production
technology is represented by a production function that relates output to inputs and the
productivity residual or shock:
Yit = F( Lit , M it , E it , K it , Ω it ),
(8)
15
which can be approximated for estimation by:
y it = β 0 + β l l it + β 2 mit + β 3 eit + β 4 k it + u it
u it = Ω it + ηit ,
(9)
(10)
where y is the log-output of plant i at time t; l, m, e and k are the log-values of labor, material,
energy and capital inputs; Ωit is the productivity shock (that is observed by the plant but not the
econometrician); and ηit is the measurement error (an unexpected productivity shock that is
unobserved by both the decision-maker and the econometrician). Thus, ηit has no effect on the
plant’s decisions, but Ωit is a state variable in the plant’s decision-making process.
Given the assumptions of the model, standard econometric models provide biased and
inconsistent estimates of equation (9) for three reasons: simultaneity between output and variable
inputs; unobserved heterogeneity in productivity; and selection bias resulting from the exit of
inefficient plants. In particular, the assumption that Ωit is seen by the plant but not the
econometrician implies that inputs are correlated with the realization of the productivity shock
(Marschak and Andrews, 1944). Plants’ higher input use resulting from a positive productivity
shock Ωit is not accounted for in the model, so OLS estimates of equation (9) will be biased
upward from this simultaneity. In addition, if profitability is positively related to K it , so a plant
with a higher capital stock expects larger future profitability at current productivity levels, it will
survive lower productivity realizations that cause small plants to exit the market. This selection
effect will cause expected future productivity to be negatively related to K it , and thus the capital
coefficient to be biased downward. Both of these impacts imply unobserved heterogeneity in
plant-level productivity shocks.
16
Unlike standard estimation methods such as OLS, the Olley and Pakes (1996) semiparametric method accounts for these issues. Applying this method first involves using the
investment decision function (7) to accommodate correlation between the error term and the
inputs, or simultaneity. This is based on the assumption that future productivity is increasing
with respect to Ωit , so plants that observe a positive productivity shock in period t will invest
more in that period, for any K it . This implies the inverse function for the unobserved shock Ωit :
Ωit = I it−1 ( I it , K it ) = g t ( I it , K it ),
(11)
which is strictly increasing in I it .
This function can therefore be used to control for the simultaneity problem. Substituting
equations (9) and (10) into (11) yields,
y it = β l l it + β m mit + β e eit + φ(i it , k it ) + η it
,
(12)
where φ(i it , k it ) = β 0 + β k k it + g t (i it , k it ) is a fourth order polynomial series estimator in
investment and capital. The partially linear equation (12) can then be estimated by the OlleyPakes semi-parametric regression method, in which estimates of production function coefficients
for variable inputs are consistent because φ controls for unobserved productivity, so the error
term is not correlated with the inputs.
Since productivity is also serially correlated, however, this does not help us to obtain a
consistent coefficient for the capital input, and, thus, to distinguish between the effects of capital
levels on investment and output decisions. Accomplishing his requires a second step to estimate
survival probabilities, to control for selection bias from the exit rule effect. This involves
applying the exit equation (6), which implies that a plant will choose to stay in the market if its
17
productivity is greater than some threshold, Ω, subject to K it . Assuming Ωit is a random walk,
ξit = Ωit − Ωit-1 , and substituting into (12), we obtain:
y it − β l l it − β m mit − β e eit = β k k it + φ t −1 - β k k it -1 + ξ it + η it ,
(13)
where φˆ t −1 results from (9), and φˆ t −1 − β k k it −1 is an estimate of Ω it −1 .
The probability of survival in period t thus depends on Ω it −1 and Ω it−1 , and thus in turn on
capital and investment at time t − 1. The probability of staying in the market for each plant is
thus calculated by a probit model, based on a polynomial series in lagged (by one period)
investment and capital stock:
y it − βˆ l l it − βˆ m m it − βˆ e e it = β k k it + g (φˆ t −1 − β k k it −1 , P̂it ) + ξ it + η it ,
(14)
where P̂it is the survival probability, and the unknown function g(·) is approximated by a fourthorder polynomial in φˆ t −1 − β k k it −1 and P̂it .
Finally, based on the estimated coefficients from (14), an unbiased estimate of total factor
productivity for the ith plant at time t, required to estimate the productivity effects of technology
transfer by comparing productivity growth for matched plants, is computed as:
ln TFPit = yit - βˆ llit − βˆ m mit − βˆ e eit − βˆ k kit .
(15)
IV. The Data
For estimation of this model, we use unbalanced panel data on plants with more than 25
employees for the Turkish apparel industry (ISIC 3222), textile industry (ISIC 3212) and Motor
Vehicle and Parts industry (ISIC 3843) from 1990-1996. Our sample represents a large fraction
of the relevant population. Textiles (manufacture of textile goods except wearing apparel, ISIC
3212) and apparel (manufacture of wearing apparel except fur and leather, ISIC 3222) are sub-
18
sectors of the textile, wearing apparel and leather industry (ISIC 32), which accounts for 35 and
20 percent of total Turkish manufacturing employment and output, respectively, nearly 23
percent of wages, and approximately 48 percent of manufactured exports. The motor vehicles
and parts industry (ISIC 3843) accounts for 5 and 10 percent of total manufacturing employment
and output, nearly 6.6 percent of wages, and approximately 5.2 percent of manufactured exports.
The data were collected by the Turkish State Institute of Statistics, from the Annual
Surveys of Manufacturing Industries, and is classified according to the International Standard
Industrial Classification (ISIC Rev.2). These plant-level data include information on the levels of
output (Y); capital (K), labor hours (L), energy (E), and material inputs (M), which are used for
the computation of the productivity variable. They also document for each year whether the plant
falls in the export status category (EXP) and the value of exports; whether the plant imported any
machinery, equipment and materials (IMP) and the value of those assets; whether the plant has
any foreign ownership (FDI); the foreign share of any imported machine and equipment in total
investment (IMPS); the share of foreign ownership (FDIS); labor shares of administrative (AWS)
and technical (TWS) workers; capital intensity (KI); export share (EXS); subcontracted input
share (SCI); and subcontracted output share (SCO), which are used as variables in the probit
regression defining the propensity score. Summary statistics of the data are reported in Table 1.
V. The Results
Our objective is to estimate the average treatment or productivity effects of foreign technology
transfer, ATT, ATE, ATU, for our data on Turkish manufacturing plants. In particular, we wish to
examine whether the performance of the treated plants is caused by foreign technology transfer
through these channels, or by employment and technological characteristics of plants regardless
of such transfers, to determine whether technology transfer is an effective way to improve plant
19
productivity. We estimate these measures alternatively for the three primary foreign technology
transfer channels identified in the literature (IMP, EXP, FDI), two productivity definitions (τ=ln
LP, and τ=ln TFP), and two matching methods (nearest neighbor, NN, and kernel). 12
Although it is difficult to directly test conditional independence, one way to assess
whether the matching approach balances the observable covariates between the group of treated
and non-treated plants is to evaluate the distribution of covariates after matching. Table 2
presents summary statistics for average levels of the (lagged) variables used in the probit
regressions on which the matching is based, for the matched and unmatched samples of
importers and non-importers, and the NN matching method.
These results show that the matching procedure was effective because there are no
significant differences found between the groups for the matched sample. For instance, the
difference in the means between importers and non-importers in terms of the logarithm of capital
intensity is 0.024 after matching (the mean logarithm of capital intensity is 2.054 and 2.030 for
the importers and non-importer after matching, respectively). This difference is 0.975 (2.0531.078) for the unmatched sample. Also, the statistically significant difference in means between
importers and non-importers before matching (t=20.574) becomes insignificant after matching
(t=0.286). Tables 3 and 4 similarly report the summary statistics of the matched and unmatched
plants comparing foreign and domestic plants, and exporting and non-exporting plants, which
exhibit similarly tight matches. 13 Furthermore, the average propensity score difference between
the treated and control groups, as illustrated in Tables 2-4, is very small (.001 for exporting and
importing and .008 for FDI). It is evident from these statistics that the balancing property of the
propensity scores is ensured by the common support restriction.
20
The average treatment effects are presented in Table 5, for both the NN and kernel
matching methods. Recall that ATT is the average treatment effect on treated plants (the impact
that foreign technology transfer through the various channels has on plants that carried out such
transfers, compared to their performance in the absence of such transfers), ATE is the effect of
transferring foreign technology on a random plant, and ATU is the average treatment effect on
the non-treated plants (the performance impact that foreign technology transfer would have had
on the plants that were not treated).
The results from the third column of Table 5 show that the average total factor
productivity (TFP) effect of importing machinery, equipment and material for plants that
transferred technology through this channel, relative to their productivity had they not been
importers (ATT), is positive and statistically significant; the estimated average effect for treated
plants was about 7.2 and 5.5 percent, for NN and kernel matching, respectively. The effect for
randomly chosen plants (ATE) is a TFP increase of 8.9 or 8.7 percent, for NN or kernel
matching. Plants that did not transfer technology through importing (ATU) similarly might have
experienced an estimated 9.4 or 9.8 percent increase in TFP had they done so.
For labor productivity the estimated effect was much greater as well as statistically
significant; the estimated average effect for treated plants was about 15 percent, for randomly
chosen plants 25-29 percent, and for non-treated plants 29-33 percent, for NN and kernel
matching, respectively, with all estimates statistically significant at the 1 percent significance
level. Overall, the results indicate a positive average TFP and LP effect of transferring foreign
technology through importing.
Further, the fifth and sixth column of Table 5 shows that the productivity of foreign
owned plants is statistically greater than the matched domestic firms. The estimated average
21
treatment effect of foreign ownership for treated plants (ATT) is an increase in TFP of about 2120 percent for the NN and kernel matching methods, respectively, which are statistically
significant at the 1 percent level for both NN and kernel matching. The ATE and ATU estimates
for randomly selected plants and those without FDI are similar in magnitude and significance,
with slightly greater effect for the kernel estimates. For example, plants that had no foreign share
would have obtained a 22 or 29 percent higher TFP had they had some foreign ownership, based
on the NN or kernel estimates. The labor productivity estimates are again much larger in
magnitude than those for TFP, with ATT estimates of 31 percent for the NN and 25 percent for
the kernel methods, and even greater estimated effects for random and non-treated plants. The
estimates overall are smaller in magnitude but similar in significance for the kernel estimations.
The last two columns of Table 5 indicate that participating in the export market also has a
significant effect on plant productivity, although the estimated impacts are smaller than for
foreign ownership, and more comparable for the different matching methods. The average effect
of exporting for plants that do export is an increase in TFP of about 12 percent, and for randomly
chosen plants is 11-12 percent. The estimated potential increase in TFP for plants that did not
export is similar in magnitude and significance, at 11-12 percent. All the measures are again
much larger, and statistically significant at the 1 percent level, for labor productivity – suggesting
increases of about 19-29 percent for exporters.
These matching method estimates confirm the suggestion of the foreign technology
transfer literature based on micro data that the primary channels through which technology
transfer has an impact are the FDI and “learning by exporting” channels – and that the former is
more significant. They also provide evidence that imports of machinery, equipment and material
have a significant productivity impact, although it is smaller. The results are also broadly
22
consistent with the Yasar and Paul (2004) study of technology transfer effects, using more
standard production function estimation, that found greater total factor productivity impacts from
FDI and exporting than importing, but still a statistically positive import impact. 14
Note also that although our analysis considers each type of foreign technology transfer
separately, there could be interactions among these effects; e.g., plants that export may also
import. This would violate the conditional independence assumption if, conditional on the
observables, other treatments predict a particular treatment. As can be seen from Tables 1 and 6,
however, the association between exports, imports, and FDI are small. Also, including each
treatment or its predicted probability as regression variables for calculating the propensity score
for other treatments did not change our findings. By contrast, the foreign technology transfer or
treatment variables are highly correlated with the licensing dummy variable, which is also
significant in regressions of treatment variables on licensing and other observables, so licensing
was not included as a treatment variable. 15
Furthermore, it is worth mentioning that when we include the lagged total factor
productivity in the probit regression to match the firms the technology transfer effects are smaller
than those presented above. Plants’ export participation as well as foreign ownership still imply
higher total factor and labor productivity, but importing machines, equipment and materials
causes higher productivity (both total factor and labor). For total factor productivity the ATE
and ATU are both positive and significant, although, ATT insignificantly positive.
VI. Concluding Remarks
Recent studies based on firm- or plant-level data have found that internationally engaged firms
are more productive than domestic ones in the same industry, even after controlling for firm or
plant size and some other observed firm or plant characteristics. What is not yet established in
23
the literature, however, is whether there are benefits associated with international activities.
More specifically, there is not enough evidence showing whether international involvement
causes a higher technical or labor productivity of firms or plants.
In this paper we attempted to shed some light on the causal effect of exporting,
importing, and FDI on the productivity of plants in three Turkish manufacturing industries
(textiles, apparel, and motor vehicles), using propensity score matching techniques, which allows
us to control for heterogeneity and selection bias in examining the relationship between
international involvement and productivity. Our results show that plants that export, import, or
have a foreign share have noticeable higher productivity even after matching the plants based on
some primary firm characteristics. This indicates that plants that are internationally involved are
not only more productive than their domestically oriented counterparts, but also that engagement
in international activities causes higher productivity.
The exporting importing, and FDI effects are similar whether one considers the imputed
effects for plants that transfer technology or the potential gain for plants that did not transfer
technology (although the former was somewhat smaller particularly for importing). Also, the
differences across alternative matching methods are not substantive, but the estimated impacts
are much greater for labor than total factor productivity. Total factor productivity is higher by
approximately 6, 10 and 25 percent for matched plants that transfer foreign technology through
importing, exporting and FDI, respectively.
Overall, these results suggest that providing incentives for firms to internalize technology
through exporting, importing, and particularly foreign ownership has the potential to improve the
productivity and thus competitiveness of Turkish manufacturing plants.
24
References
Abadie A., Drukker, D., Herr, J. L., and Imbens, G. 2001. Implementing Matching Estimators for
Average Treatment effects in Stata. The Stata Journal. 1:1-18.
Aitken, B. and Harrison, A. 1999. Do Domestic Firms benefit from Foreign Direct Investment?
Evidence from Venezuela. American Economic Review. 89:605-618.
Angris, J. 2003. Treatment Effect Heterogeneity in Theory and Practice. IZA (Institute for the Future of
Labor) Discussion Paper #851: Bonn, Germany.
Aw, B., Chung, S., and Roberts, M. 2000. Productivity and Turnover Patterns in the Export Market:
FirmLevel Evidence from Taiwan and South Korea. World Bank Economic Review, 14.
Becker, S. O. and Ichino, A. 2002. Estimation of Average Treatment Effects Based on Propensity
Scores. The Stata Journal. 2:358-377.
Bernard, A. and Jensen, J.B. 1995. Exporters, jobs and wages in US manufacturing: 1976-1987.
Brookings Papers on Economic Activity, Microeconomics 1995:67-119.
Bernard, A. and J. Wagner (1997). Exports and Success in German Manufacturing. Weltwirtschaftliches
Archiv / Review of World Economics 133:134-157.
Bernard, A. and Jensen, J.B. 1999. Exceptional exporters performance: cause, effect or both? Journal of
International Economics, 47:1-25.
Bernard, A. and J. Wagner (2001). Export Entry and Exit by German Firms. Weltwirtschaftliches
Archiv / Review of World Economics 137:105-123.
Bernard, A., J. Eaton, J. B. Jensen, and S. S. Kortum (2003). Plants and Productivity in International
Trade. American Economic Review 93:1268-1290.
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. and 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. 1986. Foreign Investment and Productive Efficiency: The Case of Mexico. Journal of
Industrial Economics. 15:97-110.
Blomström, M. and Kokko, A. 1998. Multinational Corporations and Spillovers. Journal of Economic
Surveys. 12:247-277.
Blomström M., and Wolff, E. N. 1989. Multinational Corporations and Productivity Convergence in
Mexico. NBER Working Paper. Cambridge, MA.
Blundell, R., and Costa Dias, M. 2000. Evaluation Methods for Non-Experimental Data. Fiscal Studies.
21(4):427-468.
25
Carr, D. L., Markusen, J. R. and Maskus, K. E. 2001. Estimating the Knowledge-Capital Model of the
Multinational Enterprise. American Economic Review. 91(3):693-708.
Castellani, D. 2001. Export Behavior and Productivity Growth: Evidence from Italian Manufacturing
Firms. Mimeo. ISE-Università di Urbino.
Caves, R. E. 1974. Multinational Firms, Competition and Productivity in Host Country Markets,
Economica. 41:176-93.
Clerides, S., Lach, S. and Tybout, J. 1998. Is learning by exporting important? Microdynamic evidence
from Columbia, Mexico and Morocco. Quarterly Journal of Economics. 113:903-948.
Coe, D. T., and Helpman, E. 1995. International R&D Spillovers. European Economic Review. 39:859887.
Dehejia, R. H., and Wahba, S. 2002. Propensity Score Matching Methods for Non-Experimental Causal
Studies. The Review of Economics and Statistics. 84:151-161.
Delgado, M., Fariñas, J. and Ruano, S. 2002. Firm Productivity and Export Markets: A Non-Parametric
Approach. Journal of International Economics. 57:397-422.
Doms, M., and Jensen, J.B. 1998, Comparing Wages, Skills, and Productivity between Domestically
and Foreign-Owned Manufacturing Establishments in the United States. in R. Baldwin, R. Lipsey,
and D. Richardson (eds.), Geography and Ownership as Bases for Economic Accounting. Chicago
University Press for the NBER:235-258.
Eaton, J. and Kortum, S. 1996. Trade in Ideas: Patenting and Productivity in the OECD. Journal of
International Economics. 40:251-278.
Eaton, J. and Kortum, S. 2001. Trade in Capital Goods. European EconomicReview. 45(7):1195-1235.
Girma, S., Greenaway, D., and Kneller, R. 2003. Export Market Exit and Performance Dynamics: A
Causality Analysis of Matched Firms. Economics Letters 80, 181-187.
Girma, S., Greenaway, D., and Kneller, R. 2004. Does Exporting Increase Productivity? A
Microeconometric Analysis of Matched Firms. Review of International Economics 12, 855-866.
Globerman, S. 1979. Foreign Direct Investment and Spillover Efficiency Benefits in Canadian
Manufacturing Industries. Canadian Journal of Economics. 12:42-56.
Haddad, M. and Harrison, A. 1993. Are there Positive Spillovers from Direct Foreign Investment?
Evidence from Panel Data for Morocco. Journal of Development Economics. 42:51-74.
Heckman, J.J. 1997. Instrumental Variables: A Study of Implicit Behavioral Assumptions Used in
Making Program Evaluations. Journal of Human Resources. 32:441-462.
Heckman, J.J., Ichimura, H. and Todd, P. 1997. Matching as an Econometric Evaluation Estimator:
Evidence from Evaluating a Job Training Programme. Review of Economic Studies. 64:605-654.
26
Heckman, J.J., Ichimura, H. and Todd, P. 1998. Matching as an Econometric Evaluation Estimator.
Review of Economic Studies 65:261-94.
Heckman, J.J., Ichimura, H., Smith, J. and Todd, P. 1998. Characterizing Selection Bias Using
Experimental Data. Econometrica. 66:1017-98.
Heckman, J.J., R.J. Lalonde, and J.A. Smith. 2000. The Economics and Econometrics of Active Labor
Market Programs. in Handbook of Labor Economics, Vol. III, Orley Ashenfelter and David Card,
eds. Amsterdam: North Holland:1865-2097.
Helpman, E., M. J. Melitz and S. R. Yeaple (2004). Export versus FDI with Heterogeneous Firms.
American Economic Review 94, 300-316.
Holland, Paul. 1986. Statistics and Causal Reasoning. Journal of the American Statistical Association.
81:945-960.
Keller, W. 2000. Do Trade Patterns and Technology Flows Affect Productivity Growth? World Bank
Economic Review. 14(1):17-47.
Keller, W. 2002a. Geographic Localization of International Technology Diffusion. American Economic
Review. 92:120-142.
Keller, W. 2002b. Knowledge Spillovers at the World’s Technology Frontier. Working Paper,
University of Texas, Austin.
Keller, W., and Yeaple, S.R. 2003. Multinational Enterprises, International Trade, and Productivity
Growth: A Firm Level Evidence from the United States. Working Paper, University of Texas,
Austin.
Kraay, A. 1999. Exports and Economic Performance: Evidence from a Panel of Chinese Enterprises.
Mimeo, World Bank.
Kraay, A, Soloaga, I., and Tybout, J. 2001. Product Quality, Productive Efficiency, and International
Technology Diffusion: Evidence from Plant-Level Panel Data. Presented at the NBER Summer
Meeting.
Lalonde, Robert. 1986. Evaluating the Econometric Evaluations of Training Programs. American
Economic Review. 76:604-620.
Leuven, E., and Sianesi, B. 2003. PSMATCH2: STATA module to perform full Mahalanobis and
propensity score matching, common support graphing, and covariate imbalance testing.
http://ideas.repec.org/c/boc/bocode/s432001.html. Version 1.2.3.
Meyer, B. D. 1995. Natural and Quasi-Experiments in Economics. Journal of Business and Economic
Statistics. 13:151-162.
Navaretti, G. B. and Venables, A.J. 2005. Multinational Firms in the World Economy. Princeton
University Press.
27
Olley, S. and Pakes, A. 1996. The Dynamics of Productivity in the Telecommunications Equipment
Industry. Econometrica. 64(6):1263-1298.
Rosenbaum, P. 1984. From Association to Causation in Observational Studies: The Role of Tests of
Strongly Ignorable Treatment Assignment. Journal of the American Statistical Association. 79:41-48.
Rosenbaum, P. and Rubin, D.B. 1983. The central role of the propensity score in observational studies
for causal effects. Biometrika. 70:41-55.
Saggi, K. 2002. Trade, Foreign Direct Investment, and International Technology Transfer: A Survey.
World Bank Research Observer. 17(2):191-235.
Sianesi, B. 2001. Implementing Propensity Score Matching Estimators with Stata. Prepared for UK
Stata Users group, VII Meeting, London.
Smith, J. 2000. A Critical Survey of Empirical Methods for Evaluating Active Labor Market Policies.
Swiss Journal of Economics and Statistics 136, 247-268.
Smith, J. and Todd, P. 2003. Does Matching Address Lalonde’s Critique of Nonexperimental
Estimators. Journal of Econometrics, forthcoming.
Van Biesebrock, J. 2003. Exporting Raises Productivity in Sub-Saharan African Manufacturing Plants.
NBER Working Paper, Cambridge, MA.
Vandenberghe, V. and Robin, S. 2003. Evaluating the Effectiveness of Private Education Across
Countries: A Comparison of Methods. Labour Economics, forthcoming.
Wagner, J. 2002. The causal effects of exports on firm size and labor productivity: first evidence from a
matching approach. Economics Letters. 77:287-292.
Wagner, J. 2005. Exports and Productivity: A Survey of the Evidence from Firm Level Data. Working
Paper, University of Luneburg.
Xu, B., and Wang, J. 1999. Capital Goods Trade and R&D Spillovers in the OECD. Canadian Journal
of Economics. 32:1258-1274.
Yasar, M., and Morrison Paul, C.J. 2004. International Linkages and Productivity at the Plant Level.
Manuscript.
28
Table 1. Descriptive statistics for relevant variables
(Constant Value Quantities at 1987 Prices, in ‘000 Turkish Lira)
Continuous Variables
Mean
Standard
Deviation
Minimum
Maximum
Output (Y)
Capital (K)
Labor (L)
Energy (E)
Material (M)
5,404.46
2,140.56
239.86
114.49
4022.93
34,398.31
18044.47
593.61
673.62
24,275.81
3.65
0.16
7.20
0.06
0.103
1,212,264.00
720,277.80
18141.64
22,178.53
793,684.50
Dummy Variables and Shares
Mean (Percentage of plants for dummy variables)
Log Capital Intensity (lnKI)
Import Dummy (IMP)
FDI Dummy (FDI)
Export Dummy (EXP)
IMP*FDI
IMP*EXP
FDI*EXP
Import Share (IMPS)
FDI Share (FDIS)
Export Share (EXS)
Licensing Dummy (LIC)
Subcontracted Input (SCI)
Subcontracted Output (SCO)
Administrative Share (AWS)
Technical Share (TWS)
Advertisement Expenditure Share (ADV)
Textile Industry
Apparel Industry
Motor Vehicles and Parts Industry
Small
Medium
Large
Agean Region
Black Sea Region
Central Anatolian Region
Eastern and South-East
Anatolian Region
Marmara Region
Mediterranean Region
Observations
128
23.19
4.44
57.41
2.65
18.98
3.52
20.33
2.43
22.48
4.72
11.33
14.32
15.98
2.90
0.80
9.77
68.72
21.51
50.02
21.55
28.43
20.10
1.11
5.72
1.11
69.48
2.51
7025
29
Table 2. Comparison of Importers and Non-Importers: Matched vs. Unmatched
Matched Sample
(lagged values)
Import
Log Employment
Log Real Wages
Log Capital Intensity
LIC
SCI
SCO
AWS
TWS
ADV
FDIS
EXS
4.904
5.479
2.054
0.131
0.108
0.074
0.186
0.058
0.007
0.066
0.333
Average Propensity
Score Difference
between Treated and
Control
Observations
Unmatched Sample
T-test for
the Mean
Non-Imp.
Differences
4.948
5.539
2.030
0.141
0.108
0.064
0.188
0.055
0.007
0.042
0.344
-0.664
-0.657
0.286
-0.589
0.002
0.650
-0.260
0.511
0.101
2.302
-0.398
Import
T-test for
the Mean
Non-Imp.
Differences
4.835
5.479
2.053
0.128
0.108
0.073
0.186
0.059
0.007
0.064
0.332
3.897
4.094
1.078
0.020
0.117
0.161
0.152
0.063
0.007
0.013
0.208
1341
4038
0.001
1310
763
30
35.575
33.565
20.574
16.688
-1.688
-9.039
8.555
-1.107
0.037
12.474
10.201
Table 3. Comparison of Foreigners and Domestic Plants: Matched vs. Unmatched
Matched Sample
(lagged values)
Foreign
Domestic
Log Employment
Log Real Wages
Log Capital Intensity
LIC
SCI
SCO
AWS
TWS
ADV
IMPS
EXS
5.565
6.584
2.297
0.408
0.099
0.135
0.209
0.057
0.037
0.266
0.254
5.403
6.373
2.456
0.453
0.071
0.141
0.223
0.068
0.013
0.314
0.263
Average Propensity
Score Difference
between Treated and
Control
Observations
Unmatched Sample
T-test for
the Mean
Differences
0.934
0.822
-0.800
-0.647
1.213
-0.124
-0.654
-1.493
1.028
-0.978
-0.164
Foreign
Domestic
5.461
6.584
2.297
0.408
0.099
0.135
0.209
0.062
0.037
0.266
0.254
4.059
4.333
1.274
0.029
0.115
0.139
0.158
0.057
0.006
0.063
0.238
245
5134
0.008
245
175
31
T-test for
the Mean
Differences
24.432
25.478
10.130
29.594
-1.549
-0.194
6.283
0.742
5.310
14.631
0.632
Table 4. Comparison of Exporters and Non-Exporters: Matched vs. Unmatched
Matched Sample
(lagged values)
Log Employment
Log Real Wages
Log Capital Intensity
LIC
SCI
SCO
AWS
TWS
ADV
IMPS
FDIS
Average Propensity
Score Difference
between Treated and
Control
Observations
Unmatched Sample
T-test for
the Mean
Exporters Non-Exp.
Differences
4.487
4.827
1.526
0.069
0.125
0.073
0.174
0.055
0.008
0.093
0.036
4.448
4.765
1.556
0.042
0.124
0.066
0.186
0.055
0.010
0.079
0.024
0.819
0.911
-0.379
3.295
0.104
0.455
-1.737
0.033
-0.737
1.388
1.634
T-test for
Exporters Non-Exp. the Mean
Differences
4.426
4.828
1.530
0.068
0.125
0.073
0.174
0.070
0.008
0.091
0.033
3.705
3.868
1.011
0.015
0.099
0.236
0.140
0.056
0.007
0.044
0.013
3216
2161
0.001
3139
1028
32
31.160
25.431
12.104
9.089
5.397
-19.616
9.988
4.719
0.323
7.840
5.279
Table 5. Average Effect of Importing, Foreign Direct Investment and Exporting on the
Productivity of Plants
Average Effect of
Average Effect of
Average Effect of
Importing Machines
Being a Foreign
Participating in the
and Material
Owned Plant
Export Market
ATT
Nearest
Neighbor
ATE
ATU
ATT
ATE
Kernel
ATU
Ln TFP
Ln LP
Ln TFP
Ln LP
Ln TFP
Ln LP
0.072
(0.027)*
0.089
(0.032)*
0.094
(0.038)*
0.150
(0.052)*
0.253
(0.052)*
0.287
(0.064)*
0.213
(0.059)*
0.215
(0.081)*
0.216
(0.078)*
0.314
(0.115)*
0.335
(0.118)*
0.336
(0.130)*
0.116
(0.025)*
0.112
(0.022)*
0.106
(0.033)*
0.188
(0.056)*
0.226
(0.045)*
0.283
(0.056)*
0.055
(0.017)*
0.087
(0.026)*
0.098
(0.031)*
0.154
(0.028)*
0.289
(0.039)*
0.334
(0.051)*
0.201
(0.040)*
0.287
(0.060)*
0.291
(0.063)*
0.249
(0.066)*
0.324
(0.076)*
0.329
(0.081)*
0.119
(0.020)*
0.119
(0.017)*
0.118
(0.022)*
0.285
(0.037)*
0.302
(0.034)*
0.328
(0.034)*
Notes: (1) *Significant at the 1 percent level. **Significant at the 5 percent level. ***Significant
at the 10 percent level. (2) The standard errors are bootstrapped using 500 replications.
Table 6. Correlation Coefficients for Export, Import, Foreign Ownership and Licensing
Import
Export
Foreign Ownership
Licensing
Import
Export
1.000
0.272
0.188
0.241
1.000
0.098
0.123
33
Foreign
Ownership
1.000
0.379
Licensing
1.000
Endnotes
1
See Keller (2000) and Saggi (2002) for more discussion of international technology diffusion.
2
Wagner (2002) and Girma, Greenaway, and Kneller (2003) similarly used a “nearest neighbor”
technique to examine the relationship between export status and productivity. Wagner (2002)
examined the direction of causation between productivity and exports by looking at the
performance of firms entering the export market, and Girma, Greenaway, and Kneller (2003)
determined causation through the performance of firms exiting the export market.
3
See Wagner (2005) for a survey of papers on exporting and productivity.
4
See Navaretti and Venables for the literature review of foreign direct investment.
5
See Rosenbaum and Rubin (1983).
6
See Abadie, Drukker, Herr and Imbens (2002).
7
Note that most papers in this literature use non-experimental data to examine the average
treatment effect using propensity score matching, although no dataset will include information
on all factors which might differ across observations (plants in our case). See Dehejia and Wahba
(2002) and Smith and Todd (2003) on a discussion of whether propensity score matching
generally replicates experimental results well or not.
8
See Rosenbaum and Rubin, 1983; Meyer, 1995; Sianesi, 2001; Girma, Greenaway, and
Kneller, 2003
9
We perform the matching using the psmatch2 STATA procedure developed by Leuven and
Sianesi (2003). The STATA program used for estimation of the model tests the balancing
hypothesis using an iterative process to ensure that the estimated model is consistent with this
requirement.
10
See Becker and Ichino (2002).
11
See Sianesi (2002)
12
Since non-experimental estimators can be sensitive to the specification adopted (LaLonde,
1986, and Smith, 2000), we considered the robustness of our estimators by trying alternative
arguments of the functions. For example, we ran alternative probit models to estimate propensity
scores without including the lagged levels of total factor productivity. We also estimated our
models with and without the foreign investment, export, and imported investment shares. We
found that the significance of our treatment effects were not sensitive to these changes in the
specification, which further supports our matching methods.
34
13
We also estimated the impacts of licensing technology, and found, as in most of the literature,
that the total factor productivity effect of licensing is insignificant, although an impact on labor
productivity was apparent.
14
Note also that Yasar and Paul (2004) found a significant impact of licensing on overall
productivity, which was larger than that for imports but smaller than that for FDI and exports,
whereas when the matching method was used to evaluate the impacts of licensing on TFP it was
found to be insignificant.
15
To test the robustness of our results we also defined 7 groups: EXP, FDI, IMP, EXPFDI,
EXPIMP, FDIIMP and NEITHER and carried out a multinomial probit to match the plants.
These seven treatments are mutually exclusive, and thus the conditional independence
assumption is not violated. We find even higher treatment effects with the multi-treatment
analysis. However, we view the results presented in the paper as more reliable since the number
of firms that are engaged in two international activities at the same time is very small.
35