Content uploaded by Katrin Hussinger
Author content
All content in this area was uploaded by Katrin Hussinger
Content may be subject to copyright.
Discussion Paper No. 04-12
Export Behavior and Firm Productivity
in German Manufacturing
A Firm-level Analysis
Jens Matthias Arnold and Katrin Hussinger
Discussion Paper No. 04-12
Export Behavior and Firm Productivity
in German Manufacturing
A Firm-level Analysis
Jens Matthias Arnold and Katrin Hussinger
Die Discussion Papers dienen einer möglichst schnellen Verbreitung von
neueren Forschungsarbeiten des ZEW. Die Beiträge liegen in alleiniger Verantwortung
der Autoren und stellen nicht notwendigerweise die Meinung des ZEW dar.
Discussion Papers are intended to make results of ZEW research promptly available to other
economists in order to encourage discussion and suggestions for revisions. The authors are solely
responsible for the contents which do not necessarily represent the opinion of the ZEW.
Download this ZEW Discussion Paper from our ftp server:
ftp://ftp.zew.de/pub/zew-docs/dp/dp0412.pdf
Non-Technical Summary
Why do some firms in an industry export, while others in the same
industry persistently serve the domestic market only? What are the
determinants behind these different patterns within sectors? How are
these differences in firm behavior related to productivity differences
among firms? Do the best performers go abroad, or do firms become
more productive as they serve foreign markets? This paper analyzes
these questions empirically for a sample of German manufacturing
firms.
Being currently the largest exporter of the world, the example of
Germany is of considerable interest in this context. In this paper, we
are using firm-level data from a representative survey of the German
manufacturing sector, the Mannheim Innovation Panel (MIP), to
detect the empirical relationship between firm productivity and
export status for German firms. We employ a semi-parametric
estimation approach by Olley and Pakes (1996) in order to control
for econometric problems in the estimation of firm productivities.
Using these productivity estimates, we investigate the determinants
of exporting for German manufacturing firms. High productivity
turns out to improve a firm’s odds to be an exporter, as do firms size
and R&D intensity.
The correlation between productivity and export status among firms
raises the question about the direction of causality between these two
variables. We employ the concept of Granger-causality for this
purpose, and find that there is a significant causal link from
productivity to exporting. In the opposite direction, there is no
evidence for causality. In addition, we document some descriptive
evidence about the productivity trajectory of newly exporting firms
with respect to their entry date into foreign markets. This evidence
shows that future exporters had their desirable characteristics before
entering foreign markets, supporting our causality findings.
In a last step, the paper analyzes whether exporting has a positive
impact on firm performance at all, using a matching approach. The
result of the matching analysis is that firms do not reap additional
productivity benefits from exporting. This confirms our previous
result about the direction of causality between exporting and
productivity. Summing up, this paper finds strong evidence for the
hypothesis that the best-performing domestic firms self-select
themselves into export markets, while exporting itself has no
significant impact on firm performance for German manufacturing
firms.
Export Behavior and Firm Productivity in German Manufacturing
A firm-level analysis
Jens Matthias Arnold and Katrin Hussinger1
January 2004
Abstract
This paper analyses the relationship between firm productivity and export
behavior in German manufacturing firms. We examine whether
productivity increases the probability of exporting, and assert that there
is a causal relationship from high productivity to entering foreign
markets, as postulated by the recent literature on international trade with
heterogeneous firms. In estimating productivity, we control for a possible
simultaneity bias by using semiparametric estimation techniques.
Moreover, we apply a matching technique in order to analyze whether the
presence in international markets enabled firms to achieve further
productivity improvements, without finding significant evidence for this.
We conclude that high-productivity firms self-select themselves into
export markets, while exporting itself does not play a significant role for
productivity improvements.
Keywords: Total Factor Productivity; Exports; Export-led growth; Heterogeneous firms.
JEL-Classification: F10, F13, F14, D21, L60
Addresses:
Jens Matthias Arnold
Università Bocconi
Via Salasco, 5
20136 Milano, Italy
Phone. +39-3287465190
E-Mail: jens.arnold@uni-bocconi.it
Katrin Hussinger
Centre for European Economic Research (ZEW)
P.O.Box 10 34 43
68304 Mannheim, Germany
Phone +49-621-1235-381
E-Mail: hussinger@zew.de
1 We are indebted to Giorgio Barba Navaretti, Laura Bottazzi, Dirk Czarnitzki, Christopher Flinn, Georg Licht,
Gianmarco Ottaviano and Cyrille Schwellnus for helpful comments. We also thank the team of the Mannheim Innovation
Panel (Sandra Gottschalk, Bettina Peters, Christian Rammer and Tobias Schmitt) for providing data. All remaining
errors are ours.
1
1 Introduction
Why do some firms in an industry export, while others in the same
industry persistently serve the domestic market only? What are the
determinants behind these different patterns within sectors? How are
these differences in firm behavior related to productivity differences
among firms? Do the best performers go abroad, or do firms become
more productive as they serve foreign markets? This paper analyzes
these questions empirically for a sample of German manufacturing
firms.
In response to the empirical evidence for important heterogeneity of
firms’ trade orientations within sectors in recent years, a new
theoretical strand of literature on international trade has begun to
focus on the export behavior of firms within sectors. Melitz (2004),
Melitz and Ottaviano (2003) and Bernard et al. (2003) leave behind
the assumption of a representative firm for each sector and provide
theoretical foundations for the relationship between within-sector
heterogeneity of firms and international trade in general equilibrium.
One crucial assumption of this literature is that high-productivity
firms self-select themselves into export markets. This assumption
implies a causal link from firm productivity to exporting, for which
this paper provides an empirical test.
Being currently the largest exporter of the world, the example of
Germany is of considerable interest in this context. In this paper, we
are using firm-level data from a representative survey of the German
manufacturing sector, the Mannheim Innovation Panel (MIP), to
detect the empirical relationship between firm productivity and
export status for German firms. Our data have the advantage of
achieving full geographical coverage of Germany.2 They include firms
of all size classes including a considerable number of small and
medium enterprises, and contain information about firms’ innovative
behavior. The measure for total factor productivity (TFP) used is
estimated from firm input and output data, taking into account some
econometric difficulties that arise in TFP estimation. Since firms
observe their respective productivities that are unobserved by the
researcher, they will take this knowledge into account when making
their input choices — which in turn are observed and used for the
2 Other studies that have used German data are Bernard and Wagner (1997), Bernard and Wagner
(2001) and Wagner (2002). These authors, however, use survey data from the German state of Lower
Saxony only.
2
estimation. As a result, there is likely to be a correlation between the
error terms and the explanatory variables in the estimation of the
production function, which creates a technical problem for the
estimation procedure. Least-squares estimation procedures would
produce biased coefficient estimates in this situation. Therefore, we
estimate total factor productivity at the firm level in a way that is
robust to this so-called simultaneity bias from endogenous input
choice, by using a semi-parametric estimation technique for the
production function following Olley and Pakes (1996).
Subsequently, we model the exporting decision of a firm and find
that productivity increases the odds of exporting. The positive
correlation between firm productivity and exporting that we find
does not say anything about the direction of causality: It could be
that productive firms decide to become exporters, or that exporting
makes firms more productive, or both. Trying to make a clear
distinction between correlation and causation, we employ the concept
of Granger causality to test for causal relationships in both
directions. We also document some descriptive evidence about the
productivity trajectory of newly exporting firms with respect to their
entry date into foreign markets.
Finally, our analysis goes one step further. To check the robustness
of our results regarding the direction of causality, we explicitly test
for the direction of causality opposite from the one we found using
Granger causality. To do so, we employ a matching technique, in
order to investigate whether exporting is at all effective for improving
firm performance.3 In examining this question, one has to take into
account that the subgroup of exporting firms is not a randomly
selected sample. Our previous results suggested that exporters self-
selected themselves into selling abroad because they were high
performers in the first place. To control for this sample selection
problem, our matching technique makes inferences within pairs of
firms with similar estimated a-priori probabilities of being part of the
exporting subgroup. This procedure corrects for the selection bias,
provided that the variables on which the matching process is
conditioned account for all the systematic differences relevant to both
the exporting decision and firm productivity. In other words, we
3 Using matching techniques in the context of firm exports is relatively novel. To the best of
our knowledge, only Wagner (2002) and Girma et al. (2004) have used similar methods so
far.
3
explore whether an exporting firm can reap
additional
performance
improvements from exposure to foreign markets.
There is an extensive debate on the relationship between openness
and productivity growth using aggregate, economy-wide data. Ben-
David (1993), Sachs and Warner (1995) provide empirical evidence
for a positive correlation of trade and growth. Marin (1992) finds a
causal link from exports to higher productivity growth for four
industrial countries, including Germany. Such a causal relationship
on the aggregate level can work through two channels: Either firms
become more productive as they export, or increased openness
initiates a process in which resources are re-allocated in favor of
exporting firms that are more productive than non-exporters. Our
micro-evidence that firms are unable to achieve significant
productivity gains from exporting, is evidence for re-allocation being
the primary source behind aggregate productivity gains caused by
exports.
The remainder of this paper is organized as follows: The next section
gives an overview over the related literature and the evidence
available from other countries. Subsequently, we describe our data
and give some descriptive evidence. The fourth section presents our
probit estimation results concerning the determinants of exporting
and the causal relation between firm productivity and export
behavior. In section 5, we present the results from our matching
approach, analyzing whether exporting is at all beneficial to firm
performance. Finally, the last section concludes.
2 Export behavior of firms: Where do we stand?
The statement that exporters tend to outperform non-exporters is
unlikely to cause much surprise among economists. In fact, apart
from making intuitive sense, this insight is not new. With an
increasing availability of longitudinal data at the firm level, it has
been widely documented for a number of countries, both developed
and developing. Micro-evidence on this issue is now available for the
United States (Bernard and Jensen 1999, 2001), for Chile (Pavcnik
2002), Taiwan and Korea (Aw et al. 2000), for Colombia, Mexico and
Morocco (Clerides et al. 1998), Japan (Head and Ries 2003), Spain
(Delgado et al. 2002), Italy (Castellani 2001), the German state of
4
Lower Saxony (Bernard and Wagner 2001, Wagner 2002), as well as
Thailand, Indonesia, the Philippines and Korea (Hallward-Driemeier
et al. 2002), Britain (Girma et al. 2004), China (Kraay 1999) and
sub-saharan Africa (Bigsten et al. 2002).4 The empirical literature
finds a robust positive correlation between productivity at the firm
level and exporting. The existing evidence becomes a bit thinner,
however, when asks for the direction of causality between firm
productivity and export status and thus goes beyond the analysis of
correlation, as we do in this paper.
There are at least two prominent strands of theoretical explanations
for the relationship of productivity and exporting at the firm level,
each of which emphasizes one direction of the causal relationship.
One approach has stressed the difficulties firms face in foreign
market, due to the existence of sunk costs associated to selling
abroad and fiercer competition in international markets. Roberts and
Tybout (1997), Bernard and Jensen (1999) and Bernard and Wagner
(2001) have found evidence for the existence of sunk costs in
exporting. According to this approach, above-average performers are
likely to be the ones that are able to cope with sunk costs associated
to the entry into a distant market, and make positive net profits
abroad. Also, competition could be fiercer outside the home market, a
feature that would again allow only the most productive firms to do
well abroad. This explanation is in line with the assumption made in
the theoretical literature of international trade with heterogeneous
firms that high-performing firms self-select themselves into foreign
markets. An alternative theoretical explanation for the firm-level link
between exporting and productivity puts forward learning effects
associated to exporting, implying that exporting makes firms more
productive. This view appears to be particularly prominent in the
management and policy literature. The possibility of useful
technological and managerial inputs from international contacts is
often mentioned in this context, as is the possibility of exploitation of
economies of scale by operating in several markets. As far as the
technological argument is concerned, one might expect the learning
hypothesis to have more explanatory power for countries facing
significant technological gaps vis-à-vis the foreign markets, while the
economies to scale argument may be of particular relevance for firms
from small domestic markets. Although the two explanations are not
mutually exclusive in general, the latter one shifts the burden of the
4 This list makes no claim for completeness.
5
argument onto the causal relationship from exporting to productivity,
whereas the former emphasizes the causal link from productivity to
exporting. An empirical analysis of causality is hence a means to
assess the performance of the two approaches in the data.
One of the first studies to make a clear empirical distinction between
correlation and causality is Bernard and Jensen (1999). They find
that exporters have all their desirable characteristics before taking up
exporting, and that the performance paths of exporters and non-
exporters do not diverge following the launch of export activities by
the former. Using a slightly different methodology, Clerides et al.
(1998) also find strong evidence for self-selection in their data from
Colombia, Mexico and Morocco. They do not find any evidence for
learning effects from exporting. For Taiwan, Aw et al. (2000) find
that newly exporting firms outperform other firms before entry, and
in some industries they experience productivity improvements
following entry. Continuous exporters do not increase their
productivity advantage vis-à-vis non-exporting firms over time. These
results are consistent with the self-selection hypothesis, and lend only
limited support to the learning hypothesis. For Korea, the correlation
between export status and firm productivity is less crisp, but they
find no support for the learning hypothesis here. Delgado et al.
(2002) apply non-parametric methods on a panel of Spanish firms.
Their results support the self-selection mechanism of highly
productive firms into exporting, while the evidence for learning
effects is not significant. Only when limiting their sample to young
firms do they find some evidence for learning effects. On the other
hand, Kraay (1999) and Bigsten et al. (2002) find evidence for
learning effects for China and several Sub-Saharan African countries,
respectively. Castellani (2001) finds that Italian firms with a very
high exposure to foreign markets experience learning effects, while
below this threshold export intensity this is not the case. In the
remainder of this paper, we look for evidence both for the self-
selection hypothesis and the learning hypothesis in German data.
3 Data and Descriptive Statistics
The underlying database is an extract from the Mannheim
Innovation Panel (MIP), conducted by the Centre or European
Economic Research (ZEW) on behalf of the German Federal Ministry
6
for Education and Research (BMBF). With its principal focus on
firms’ innovation behavior, the MIP is the German part of the
Community Innovation Survey (CIS) of the European Commission.
Started in 1992, the representative survey collects yearly information
from firms in the manufacturing sector all over the country. The
survey includes firms of all size classes, including a large number of
small and medium firms that are not obliged to publish their
accounts by German law. This study uses an unbalanced panel of
2,149 observations on the firm level in the years from 1992 to 2000.
On average, there are 5.52 years of data per firm available. Our data
have the advantage of achieving full geographical coverage of
Germany, including West and former East Germany. A drawback of
our data set is its relatively limited size, which restricts us in our
choice of methodology.5
The data contain information on the export value of each firm. We
consider as exporters those firms that sell more than a threshold
value of 5% of their turnover abroad. In the light of Germany being a
highly open economy in an increasingly integrated Europe, we
consider this definition adequate for the sake of identifying those
firms as exporters that have a minimum interest in their activities
abroad. By using this definition, we want to abstract from minimal
trade relationships due to border proximity and focus instead on
systematic and significant foreign sales activities. 1,260 observations
belong to exporting firms according to our definition. This
corresponds to 227 firms in the sample that conduct exports in every
observed year, whereas 112 firms have no exports in any sample year.
Table 1 shows descriptive statistics for exporting and non-exporting
firms.
The first step of our analysis is to arrive at an appropriate estimate
of total factor productivity (henceforth TFP) at the level of the firm.
Productivity is unobservable and has to be estimated using
observable factor inputs and outputs. We assume a two-factor Cobb-
Douglas production function containing labor and capital, and
construct our TFP measure from the residual of each observation in
the logarithmic form of the equation. However, there is a technical
caveat in this estimation procedure. Using ordinary least squares
methods to estimate the factor coefficients is likely to produce biased
5 As an example, applying a GMM approach following Blundell and Bond (2000) is not
possible with our data.
7
estimates, due to a correlation between the exogenous variables and
the error term in the logarithmic estimation equation. The
productivity of a firm -which is unobserved by the econometrician
and represented by the error term in the estimation equation- is
expected to influence the factor input decision, the outcome of which
are the observed input factors on the right hand side of the equation.
This econometric problem is commonly known as the simultaneity
bias, first mentioned by Marschak and Andrews (1944).
Therefore, in line with previous studies such as Bernard and Jensen
(2001a) and Pavcnik (2002), we employ a semi-parametric estimation
technique following Olley and Pakes (1996) to get consistent
estimates of TFP. This estimation method produces factor coefficient
estimates that are robust to the presence of simultaneity and
unobserved heterogeneity in production, without significantly
increasing the computational burden.6 Appendix A briefly outlines
our estimation procedure for TFP. The limited size of our sample
requires us to estimate the production function on a relatively high
level of aggregation, dividing the manufacturing sector into four
separate industries. Details of this aggregation are found in Appendix
B. For the remainder of the paper, we use productivity as a relative
measure, dividing it over the average level in the same year and
industry at the NACE2-level. This specification allows us to focus on
firm heterogeneity within sectors.
A comparison of our TFP estimates between exporters and non-
exporters reveals important exporter premia in terms of productivity.
In addition to our TFP estimates, our analysis uses firm size, R&D
behavior and wages as well as firms’ location (East or West
Germany) as explanatory variables. Exporters and non-exporting
firms display notable differences in those characteristics. Exporting
firms are larger than non-exporting firms. On average, they have
almost three times as many employees, and approximately the same
holds for turnover. In our subsequent regressions, we use the log of
the number of employees to account for firm size, because of the
skewed size distribution of firms in our sample.
6 The data contain no information as to whether a firm that exited the sample also left the
market or not. Thus, it was not possible to control for a possible selection bias caused by
non-random patterns in the exit of firms from our sample, although the methodology used
would in principle allow for this.
8
Table 1: Descriptive Statistics of Exporters vs. Non-exporters
Variable Ex
p
orters Non-Ex
p
orters
N=1,260 N=889
TFP 1.51 1.10
TFP relative to average in industry and year 1.09 0.82
Export intensity 0.35 -
Number of employees 330 116
Sales in millions of Euro 96.89 27.64
Innovator (yes/no) 0.54 0.26
R&D expenditure in mio. Euro (if innovator) 3.64 0.54
R&D intensity (if innovator) 0.04 0.06
Share of sales from new products 4.69 2.58
Wage per employee 66.27 53.15
Age 40.01 26.96
East Germany 0.22 0.50
A particular advantage of our data set is that we have information
on the innovative efforts of firms, which allows us to use two
variables related to innovation. We include these variables to control
for the importance of technology for trade flows at the firm level. Our
first measure is firm expenditures in research and development. The
share of firms that invest in R&D is about two times higher among
the exporting firms in our sample (see table 1). The bulk of this
expenditure occurs among exporting firms. Looking at R&D
intensities defined as R&D expenditures as a fraction of turnover,
however, reverses this picture, with the average R&D intensity being
lower for exporting firms. Another variable we use is the percentage
of sales that originate from products newly introduced to the market.
This variable controls aspects of the product innovation activities like
marketing costs that are not captured by R&D expenditures. An
obvious caveat with this variable is that the definition of a new
product is at the discretion of the firm itself. Having a new product
may encourage a firm to expand into foreign markets. Bernard and
Wagner (1997) and Bernard and Jensen (2001) use a binary variable
for the introduction of new products. We prefer to use the share of
sales of new products instead, on the basis that this may be a more
appropriate indicator for the value of the new product to the firm.
This share is considerably higher for exporting firms.
9
In addition, we include the average wage defined as the total wage
bill divided by the number of employees. This wage proxy is the only
information that we have about skill composition of a firm’s labor
force. In competitive factor markets, the quality of labor is positively
related to the wage. At the same time, however, TFP also as a
positive influence on wages, and we are unable to disentangle the two
effects on wages. In our sample, exporting firms pay higher average
wages, suggesting an extended use of skilled labor among exporters.
The particular situation of Germany with its turbulent recent history
calls for the inclusion of a dummy variable for the formerly socialist
part of the country. Since the 1989 fall of the Berlin wall, East
Germany has been undergoing a transition process from a planned
economy into a market economy. Several empirical investigations
indicate that the transition process has not concluded yet.7 A dummy
for East German firms captures the differences caused by firm
location. Table 1 shows that the group of non-exporting firms
contains more than twice as many East German firms as the group of
exporters.
Finally, the data contain information on the firm age. Generally, firm
age has the problem of being correlated with several other variables
we use, such as size, wages and productivity. Moreover, a firm may
have undergone ownership changes, implying that the concept of
continuity that one would suppose behind firm age may be badly
represented by this variable, particularly at the upper end of the age
distribution. Also, a firm is unlikely to gain more experience once it
has reached a certain threshold age. For relatively young firms,
however, age may be important. This is why we use age as a binary
variable indicating the lower third of the age distribution, situated at
approximately 10 years of age. We return to this issue in the
discussion of our regression results in the next section.
4 What characterizes an exporting firm?
The next step of our analysis is to identify those firm characteristics
that make a firm more likely to export. In other words, we are
interested in the dividing line between firms that sell only
domestically and those that export to foreign markets. Our
7 See Czarnitzki (2003) as an example.
10
theoretical model behind the export decision of a firm is
straightforward. In the absence of sunk costs, a rational profit-
maximizing firm exports if the current expected revenues from
foreign sales exceed the cost of production and shipping for the
foreign market. Whether or not this is the case for an individual firm
is assumed to depend, among other things, on a vector of firm-
specific characteristics X. In any period, a firm will export whenever
exporting carries an additional positive net profit:
0)1(),( 1>−⋅−− −
itititititit YSqXcqp for the foreign market,
where p is the export price, q the exported quantity, c are additional
production costs of producing q, S are sunk costs of exporting and Y
is a binary variable indicating whether a firm exports or not.
If there are sunk costs involved in taking up export activities, a
dynamically maximizing firm will look beyond the present period
when deciding whether to export. The presence of sunk costs makes
the decision rule dynamic, because exporting today carries an
additional option value of being able to export tomorrow without
paying the sunk costs of exporting. The value function of this
dynamic problem can be expressed as:
{}
()
()
11
max ( , ) (1 )
0, 1
it it it it it it it it
t
VpqcXqSYEV
Yδ
−+
=−−⋅−+⋅
∈
where delta is a discount factor. The solution to this problem is the
decision rule
[]
î
í
ì>=−=⋅+−
=++
otherwise
YVEYVEqXcqp
Yititititititititit
it :0
0)0|()1|(),(:1 11
δ
.
The last term of this expression represents the option value of
exporting. In this decision rule, the firm- and time-specific
realizations of the vector X determine different decision outcomes
across firms and time. In other words, we are explaining different
export decisions by firms with observation-specific firm
characteristics. Particularly, we are interested in the effect of firm
productivity as one element of that vector. If the option value due to
sunk costs is indeed taken into account in the decision, we should
also expect lagged values of the dependent variable to have
11
explanatory power in the empirical implementation of this model. In
order to estimate the export decision, we translate the theoretical
model into an empirical probit model in which export behavior
depends on a variety of observed, firm-specific characteristics:
P(Y
it
=1)=
Φ
(TFP
t-1
, size
t-1
, RD
t-1
, NP
t-1
, skills
t-1
, east, young, D
it
)
where Φ is a normal cumulative density function, TFP is our
estimated (relative) total factor productivity, size is proxied by the
logarithm of employees, RD are expenditures in research and
development as a fraction of turnover, NP captures the introduction
of new products by a firm as explained in section 3, skills are proxied
by average wages, east is a dummy for the formerly East German
states and young is proxying age in the form of a binary variable
indicating the lower third of the age distribution. All variables on the
right hand side are lagged one period. Finally, we also include
dummy variables for the sector and the year of observation to
capture time- and industry-specific effects not specific to an
individual firm.8 Bootstrapped standard errors are used to test the
significance of the coefficients. We are estimating two different
specifications of the above equation. First, we take our entire sample
in the first column of table 2. In a second glance, we look only at the
subsample of firms that do not switch export status and abstract
from the lagged dependent variable to check for the robustness of our
previous results.
The estimation results for the whole sample identify several variables
with significant explanatory power for the export decision. Sunk costs
are a key determinant of the export decisions for the firms in our
sample. In quantitative terms, this effect is very large: A discrete
change from zero to one in the lagged export status increases the
estimated probability of exporting by 80%, at the means of all
remaining variables. These results are in line with the findings in
Roberts and Tybout (1998) and Bernard and Wagner (2001).
Another variable with a significant positive influence on the export
decision is, as expected, firm productivity. The coefficient is positive
and different from zero at a confidence level of 93%, implying that
8 Due to the limitations of our data, the industry dummies have to be highly aggregated. We
use four different industry dummies for the manufacturing sector each year. See appendix B
for details on the aggregation used.
12
high-productivity firms are significantly more likely to be exporters.
A larger firm size also makes a firm more likely to export. Moreover,
the effort a firm puts into R&D increases the odds of exporting, while
the same does not hold for the share of new products in this
specification of the model. Hence, one of our innovation variables has
significant explanatory power for the export behavior of firms here.
Firms located in the East of Germany are significantly less likely to
export, suggesting that they are still lagging behind with respect to
competitiveness in international markets. The quantitative effect of
location is considerable: At the means of all other variables, location
in the East reduces the probability of exporting by almost 12
percentage points. Even for a firm with high productivity, the
negative impact of location in the East hardly diminishes.
Table 2: Probability of Exporting
Probit Estimates Complete Sample Only non-switchers
Dependent Variable: Export Status N=2,037 N=1,369
TFP 0.15*
(1.84)
0.25***
(2.60)
Lagged Export Status 2.61***
(29.89)
-
Size (log of employment) 0.12***
(3.73)
0.53***
(14.68)
R&D-Intensity 2.01***
(2.79)
11.27***
(6.65)
New Product Share 0.003
(0.78)
0.008*
(1.84)
Average wage 0.91
(0.37)
5.52**
(2.22)
East Germany -0.31**
(-1.96)
-1.09***
(-6.40)
Young 0.24
(1.58)
0.35**
(2.16)
Year Dummies Included. Included.
Industry Dummies Included. Included.
Pseudo-R2 0.61 0.38
All explanatory variables are lagged one year.
Z-values in parentheses, based on bootstrapped standard errors.
*, **, *** indicate statistical significance at the 10%, 5% and 1% level, respectively.
13
In a second specification of our probit model, documented in the
second column of table 2, we repeat the estimation for only those
firms with persistent export behavior in our sample, which excludes
the lagged dependent variable from the set of regressors. We are
aware of the fact that this is a somehow arbitrary selection, since
firms that we observe as non-switchers of export status may indeed
switch inside our time window. Restricting our attention to this
subsample, however, enables us to abstract from the effect of sunk
costs. As it turns out that past exporting has a remarkably strong
explanatory power for the current realization of the export status,
this selective specification allows us to check for the robustness of the
effects of the remaining explanatory variables in our model.
The results from this specification are qualitatively very similar to
the previous ones, with generally higher levels of statistical
significance of the coefficient estimates. Again, productivity
significantly increases the odds of exporting, as do firm size and R&D
intensity. The share of new products in a firm’s product portfolio is
now a significant predictor of the export status, with a positive effect
on exporting. Moreover, the model predicts higher chances of
exporting for firms with high-skilled employees, proxied by a high
average wage. We are aware of the fact that our proxy is not a
perfect one, since it is likely to be correlated with TFP, but we do
not avail of any better proxy for skills. Concerned about the
correlation between two of our regressors, we ran the estimation
without the wage-variable, and found the results very similar to the
ones reported in table 2.
As for the complete sample, our estimation suggests that firms
located in the formerly socialist part of Germany are significantly less
likely to export.
Finally, we are using age as a binary variable indicating the lower
third of the age distribution. This formulation is due to several
reasons: We are concerned about a correlation of age with several
other variables in the regression, such as firm size, wages or
productivity. Moreover, while we do observe age, we do not observe
whether there has been continuity in ownership or management over
a firm’s lifespan. Some of the firms in our sample are aged well above
100 years, and it is doubtful whether age conveys any relevant
information for the export decision at this high end of the
distribution. On the other hand, for young firms age may well have a
14
relevant influence. Therefore, we use a binary variable for the lowest
third of the age distribution, which turns out to be 10 years.
We interpret the positive coefficient as suggesting the existence of
some firms that were founded with an immediate focus beyond the
domestic market. It could be the case that this result reflects the
increasing degree of European trade integration at the end of the
twentieth century, culminating in the 1992 Maastricht Treaty. Due
to the large amount of turbulence in East German manufacturing
following the German reunification, there is a disproportionate share
of young firms in East Germany. Still, our coefficient estimates
display opposite signs for the respective binary variables indicating
young firms and East German firms. This suggests that our firm age
specification indeed captures an independent influence of age on the
firm export decision. Age turned out to be insignificant in any other
form (linear, quadratic, or other dummy and spline combinations).
We retain as one key result from the model of the export decision
that more productive firms are more likely to be exporters. Having
ascertained this, we are now interested in the direction of causality
between the two variables. As a first glance, we document some
descriptive evidence of the relationship between firm productivity
and export status across the time dimension. For this purpose, we
have singled out the firms that initiated export activities during the
time frame of observation. Figure 1 depicts as a bold line the
trajectory of the relative productivity measures of these firms (with
respect to the average in the same year and NACE2-sector). Each of
them took up exporting at time t, which of course represents
different years across the observations. As a means of comparison,
the figure also depicts (as a dotted line) the average productivity of
firms that persistently serve the home market only.
At time t-3, the future export starters are part of the group of non-
exporters, although we know that they will emerge from this group
and take up exports in three years to come. Their average
productivity at t-3 is almost equal to the one of those firms that will
not take up exporting later on. In the two periods preceding the
export market entry, future exporters experience a significant
increase in TFP, but this tendency does not continue after export
market entry. Once they are exporters, these firms continue to have
an average productivity above the average TFP of continuous non-
exporters, but the productivity gap with respect to the latter does
15
not widen any further, and the growth tendency is not maintained.
Unfortunately, the limited size of our data does not allow us to make
formal inferences between the two subgroups depicted in figure 1.9
Still, we interpret these patterns as descriptive evidence that our new
exporters may well have taken their initial export decision in reaction
to their performance trajectory, while it is unlikely that their TFP
benefited largely from the export decision itself.10
In order to make a formal test of the causal relationship between
productivity and exporting, we use the concept of Granger causation:
A variable X is said to granger-cause a variable Y if lagged values of
X can help to predict current values of Y significantly better than
own lagged values of Y. For this reason, we estimate two separate
vector auto-regressions of productivity and exporting, using fixed
effects to capture unobserved heterogeneity among firms:
9 Such a comparison is the basic approach for causal inference in related studies such as
Bernard and Jensen (1999), Clerides et al. (1998), or Aw et al. (2000).
10 It seems remarkable that firms actually loose some of their productivity advantage as they
take up export activities. The reasons behind this fact could be an interesting topic for
further research, although our data do not allow us to go much deeper on this observation.
Figure 1: TFP Trajectory of New Exporters
0.7
0.8
0.9
1
1.1
1.2
1.3
-3 -2 -1 0 1 2 3
Time (Export market entry in t=0) New Exporters
Firms that never export
16
åå
=
−
=
−++⋅+⋅=
2
1
111
2
1
1
jtijitj
jjitjit YTFPTFP
εκγβ
åå
=
−
=
−++⋅+⋅=
2
1
222
2
1
2
jtijitj
jjitjit YTFPY
εκγβ
In other words, we estimate a linear model of the influence of lagged
values of productivity and export status on current firm productivity,
allowing for firm-specific means, and a linear probability model of the
export status on its lagged values and those of productivity, allowing
again for firm-specific means. Since our descriptive evidence in figure
1 suggests that most of the movement in the productivity trajectory
of firms takes place in the two periods preceding export market
entry, the use of two lags in the VAR estimation appeared to be the
most obvious choice here. Due to the heteroscedasticity present in
linear probability models, we use Huber/White/Sandwich robust
standard errors in both equations. Subsequently, we perform Wald-
tests to test the joint significance of the coefficients of the two lagged
values of the variable that is not on the left hand side of the
respective regression.
As shown in table 3, the lagged values of productivity have
significant explanatory power for predicting current export status;
the coefficients are jointly significant at the 5%-level. On the other
hand, lagged values of the export status do not have significant
explanatory power for predicting current productivity at any
conventional level of statistical significance. This leads us to the
conclusion that productivity granger-causes exporting in our data,
while the opposite is not true.11
Table 3 : Testing for Granger Causation
Dependent Variable Null hypothesis F-Statistic
TFPt
(Current Productivity)
(1) Yt-1=0
(2) Yt-2=0
F(2,1235) = 0.28
Prob > F = 0.75
Yt
(Current export status)
(1) TFPt-1=0
(2) TFPt-2=0
F(2,1312) = 3.12
Prob > F = 0.04
11 In statistically correct language, our results imply that we cannot exclude Granger non-
causation from exporting to productivity, while we can exclude non-causality from
productivity to exporting at a confidence level of 95%.
17
We have checked this result for robustness to the specification of
variables used here. In particular, we have used formulations with
two continuous variables (export intensity and productivity), with
two binary variables (above average productivity and export status),
and used conditional logit models with fixed effects instead of linear
regression models where the dependent variable was binary. We have
also used the absolute estimates of productivity instead of the
relative ones we use throughout the paper, and changed the number
of lags to one or three. The qualitative results remain unchanged
throughout.
5 Does Exporting improve productivity at all?
The results from the preceding section speak quite a clear language:
Our data exhibit a causal relationship from firm productivity to
export status in the Granger sense. In order to check the robustness
of this result, this section turns the perspective around and looks for
a causal link working in the opposite way. We are now interested in
examining whether there is any causal relationship at all from
exporting towards productivity that we may not have detected with
the method applied above. If our previous results are robust, we
should not be able to detect such a causal link. This section employs
a matching technique to make consistent comparisons between
exporters and non-exporters in our sample, regarding TFP in levels
and growth rates. Our aim is to assess the causal effects of a
treatment, exporting, on the treatment group, the exporting firms.12
This setup bears close resemblance to situations encountered in the
microeconometric evaluation of active labor market policies, as
surveyed in Heckman et al. (1999).13 In that literature, the research
interest lies in identifying the causal effect of a treatment, which
could be a training program. The natural variable of interest for the
evaluation of the treatment is the difference between the average of
an outcome variable of a treatment group that participated in a
program, and the average outcome variable in the counterfactual
situation of that same group not having participated. The problem is
12 See Rosenbaum and Rubin (1983) and Heckman et al. (1998) for a more comprehensive
discussion of matching methods.
13 Matching Methods have also been applied in other contexts, such as the effects of R&D
subsidies on firms, e.g. Almus and Czarnitzki (2003).
18
that by definition, the latter case is not observed. Comparing simple
averages of a treatment group and a control group, however,
produces biased results, because the selection mechanism that
governs entry into the treatment group is a non-random process.
Matching methods offer a solution to this “missing data problem” by
undertaking comparisons between the average outcomes of a
treatment and a control group conditional on a vector of observable
variables X instead, where X is assumed to influence the selection
decision. Each element of the treatment group is appropriately
matched with one (or more) elements of the control group. In this
conditional sample, one can then assume that elements of both
groups exhibit no systematic differences relevant to the selection
process, a statement that can not be made unconditionally. Hence,
while there is no control element with which one could compare a
treated element unconditionally, matching techniques assume that
one can undertake such comparisons conditional on the observed
realizations of X. All comparisons are hence made within the
matched pairs, and the effects of treatment averaged over all
elements of the treatment group. The so-calculated effect of the
treatment variable is called the average treatment effect on the
treated, and can be given a causal interpretation.
Of course, applying a matching technique requires that one can
correctly identify the determinants of selection into the treatment
group, which are the exporting firms in our sample. The empirical
model of the export decision estimated in section 4 is able to classify
correctly 92% of the observations into their respective export status.14
This gives us confidence that we have identified an appropriate
mapping from the observed firm characteristics into the export
status. In other words, we dispose of an appropriate model for the
selection mechanism to apply matching methods.
A crucial assumption for the validity of applying matching is the
assumption of conditional independence. This assumption is satisfied
as long as the fact that one firm takes up export activities does not
affect the outcome variable (productivity) of the non-exporting firms.
The result of firm productivity driving own export status and not
vice versa in section 4, suggests that firm productivity is not very
14 Of 2,037 observations, 72 were incorrectly predicted to be exporters, while 94 were
wrongly predicted to serve the domestic market only. Hence our prediction errors are more
or less balanced between the two types of errors possible.
19
sensitive to own export status (the verification of which is our aim in
this section), and it should be even less likely to react to the export
status of other firms in the sample. Moreover, the data exhibit a
persistent coexistence of exporting and non-exporting firms in the
same sectors, and despite a notable amount of turbulence between
these two groups, there exporters display a persistently higher
productivity. Hence there is no reason to believe that the conditional
independence assumption is violated in our case.
Our matching technique is one-to-one, i.e. it undertakes comparisons
within pairs of observations, conditional on a vector X.15 The
variables contained in this vector are the explanatory variables used
the probit model of section 4, for the whole sample. Each exporting
firm is thus matched with one non-exporting firm in a manner that
minimizes the within-pair difference in the estimated probability of
having taken up exports (the so-called propensity score). In addition
to the propensity score, we decided to take firm size and location in
East or West into account in creating the matched pairs, in order to
guarantee some minimum level of homogeneity within our matches.16
The matching is implemented in Stata 8 using the psmatch2
procedure by Leuven and Sianesi (2003).
The matching procedure has been able to assign a match to all but
30 of the exporting firms. This is the case because we prefer a
cautious formulation by not assigning a match to exporters with a
higher propensity score than the highest one of a non-exporting firm
to satisfy the common-support condition. A total number of 840 non-
exporting firms have been assigned as matches to 1,167 firms, where
a control observation can be assigned more than once in the
matching process. The within-pair differences of the propensity score
are quite small, with an average of 0.005 and a standard deviation of
0.043. This suggests that our matching process has been able to find
appropriate matches.
Table 4 shows the averages on the outcome variables productivity
and its growth rates for exporters (the treated) and non-exporters
(the controls) in the first two columns. The third column contains
15 We used a t-test to infer whether the distances to the nearest neighbors in both directions
are symmetrical, in order to assure that matching with only one nearest neighbor does not
introduce a bias. For 99.99% of the treatment observations, symmetry could not be rejected
at the 1% significance level.
16 The distance measure used to condition on the three variables is Mahalanobis distance.
20
the average difference of the outcome variable between these two
groups for the unmatched sample. This is the same result obtained in
table 1, i.e. a simple mean comparison between exporters and non-
exporters. Looking at TFP in levels, we find that for the unmatched
sample, exporters are on average more productive by about a quarter
of the average TFP in each sector and year. Once one considers the
inference within the matched pairs, however, this difference becomes
very small, as can be seen in the rightmost column of table 4. This
difference within the matched pairs is called the average treatment
effect on the treated (ATT), and is the interesting result for a causal
interpretation.
Table 4 : Matching Results
Treated Controls Diff. of sample
means
ATT
(Std.Dev.)
Outcome Variable: TFP
N=2,037
Unmatched Sample N=1,197 N=840
1.09 0.81 0.27
Matched Sample N=1,167 N=840
1.07 1.04 0.03
(0.04)
Outcome Variable: TFP growth 1 year later
N=1,170
Unmatched Sample N=706 N=464
.089 0.11 -0.02
Matched Sample N=677 N=464
.089 0.10 -0.01
(0.09)
Outcome Variable: Cumulative TFP growth 2 years later
N=1,170
Unmatched Sample N=706 N=464
0.14 0.16 -0.02
Matched Sample N=677 N=464
0.13 0.15 -0.01
(0.04)
In other words, as one controls for the selection bias of the treatment
group, the productivity differences between the correctly chosen
objects of comparison decrease notably in our data. In order to assess
the statistical significance of this remaining positive difference, we
use bootstrapped standard errors. These are reported below the
21
average treatment effects. Comparing the average treatment effect on
the treated of approximately 0.03 with our bootstrapped standard
error of approximately 0.04 shows that while the difference is
positive, it is not significantly different from zero at any conventional
level of statistical significance. Hence we conclude that once we
control for the bias induced by the non-random sample selection,
there are no more significant productivity advantages for exporters.
Looking at productivity growth instead of levels, we find that the
average TFP growth of exporters is slightly slower than for non-
exporting firms.17 This holds whether we define the growth rates over
a time frame of one or of two years ahead from the observation time.
In other words, once a firm is an exporter, its productivity does not
grow faster on average than that of an average non-exporting firm,
regardless of whether one applies matching or not. Again,
bootstrapped standard errors reveal that the difference is statistically
insignificant. Note, however, that exporters have a higher average
TFP level than non-exporting firms.
Summing up the results from the application of the matching
procedure, we find that once we control appropriately for the
selection into the treatment group, there are no significant causal
effects from exporting towards TFP, neither in levels nor in growth
rates over one or two years following the observation date. The
results from the Granger causality tests in section 4 are thus
confirmed by the results of the matching analysis.
6 Conclusions
In this paper, we have examined the relationship between export
behavior and total factor productivity at the firm level, using a
representative sample of German manufacturing firms. Firm
productivities are estimated using a semiparametric estimation
method following Olley and Pakes (1996). We find that those firms
that serve foreign markets are above average performers in terms of
productivity. In our model of the export decision of the firm,
productivity increases the probability of exporting.
17 When examining growth rates of productivity, we refer to growth rates of absolute TFP
rather than the relative measure we use throughout the rest of the paper. The results are
qualitatively similar, however, for both TFP measures.
22
In order to determine the direction of causality between exporting
and productivity, we estimate vector auto-regression models with
fixed effects for the two variables, and run Granger-causation test in
both directions. We find that exporting does not Granger-cause
productivity, while in the opposite direction we do detect a causal
relationship in the Granger sense. We also depict the productivity
trajectory of future export starters with respect to their entry date
into foreign markets, and find that these firms tend to have their
desirable performance characteristics already before taking up export
activities. These results suggest that the direction of causality runs
from productivity to exporting, and not vice versa.
Finally, we go one step further and explicitly test for productivity
gains from exporting. We use our empirical model of the export
decision to predict the probability of a positive export decision for
the firms in our sample. Then we compare the productivities between
exporters and non-exporters, conditional on the estimated
probabilities of exporting, as well as on size and on geographical
location (East or West Germany). We make inferences within
matched pairs of exporters and non-exporters. By employing the
matching method, we control for the non-random selection of
exporting firms in our sample, and interpret our results as causal. We
find no significant productivity differences between exporting and
non-exporting firms within the matched pairs, neither in levels nor
growth rates, and conclude that there are no statistically significant
productivity gains from exporting in our sample.
Our results concerning the direction of causality can hence be seen as
quite robust: Causality runs from productivity to exporting, and not
vice versa. The good ones go abroad, while exporting itself does not
help a firm to improve its productivity. This result supports the
selection mechanism assumed in recent theoretical models of
international trade with heterogeneous firms (Melitz 2004, Melitz and
Ottaviano 2003, Bernard et al 2002). In these models, intra-sectoral
differences in export behavior are explained by exogenously different
productivity levels of firms, with the high-productivity firms serving
foreign markets. According to the results of our analysis, this
assumption seems appropriate for the case of German manufacturing.
From an industrial policy perspective, there is hence no reason why
German policy makers should prefer foreign sales over domestic sales.
23
Where policy aims at creating new exporters that have not to date
been exceptional performers, one has reason to wonder whether such
firms will ever be able to survive in international markets without
public support. Our results show no support for the hypothesis that
firms will become better performers once they are active in foreign
markets. Given the fact that Germany is generally considered a
technologically advanced economy with a significant domestic market
size, these results may be different for firms from other economies,
where technological spillovers from exporting or economies of scale
are more likely to matter.
24
References
Almus, M. and Czarnitzki, D. (2003). The Effects of Public R&D Subsidies on
Firms’ Innovation Activities: The Case of Eastern Germany.
Journal of
Business and Economic Statistics
21(2), 226-36.
Aw, B. Y., Chung, S. and Roberts, M. (2000). Productivity and Turnover in the
Export Market: Micro Evidence from Taiwan and South Korea.
World
Bank Economic Review
14, 65-90.
Ben-David, D. (1993). Equalizing Exchange: Trade Liberalization and Income
Convergence.
Quarterly Journal of Economics
108, 653-79.
Bernard, A. B. and Jensen, B. (1999). Exceptional Exporter Performance: Cause,
Effect, or Both?.
Journal of International Economics
47, 1-25.
Bernard, A.B. and Jensen, B. (2001). Why some Firms Export. NBER Working
Paper 8349.
Forthcoming in The Review of Economics and Statistics
.
Bernard, A.B. and Jensen, B. (2001a).
Exporting and Productivity: The
Importance of Reallocation
. Mimeo, Dartmouth College.
Bernard, A.B. and Wagner, J. (1997): Exports and Success in German
Manufacturing.
Weltwirtschaftliches Archiv
133, 134-57.
Bernard, A.B. and Wagner, J. (2001). Export Entry and Exit by German Firms.
Weltwirtschaftliches Archiv
137, 105-23.
Bernard, A.B., Eaton, J., Jensen, B. and Kortum, S. (2003) Plants and
Productivity in International Trade.
American Economic Review
93, 1268-
90.
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.
Castellani, D. (2001).
Export Behavior and Productivity Growth: Evidence from
Italian Manufacturing Firms
. Mimeo, ISE-Università di Urbino.
Clerides, S.K, Lach, S. and Tybout, J. (1998). Is Learning-by-Exporting
Important? Micro-Dynamic Evidence from Colombia, Morocco, and
Mexico.
Quarterly Journal of Economics
113(3), 903-947.
Czarnitzki, D. (2003),
Das Innovationsverhalten von Unternehmen und die Rolle
der Forschungs- und Technologiepolitik, Ein Vergleich zwischen Ost- und
Westdeutschland
, doctoral thesis, Essen.
Delgado, M., Fariñas, J.C., and Ruano, S. (2002): Firm Productivity and Export
Markets: A Nonparametric Approach.
Journal of International Economics
57, 397-422.
Girma, S., Greenaway, D. and Kneller, R. (2004). Does Exporting Lead to Better
Performance? A Microeconometric Analysis of Matched Firms.
Forthcoming in Review of International Economics
.
25
Hallward-Driemeier, M., Iarossi, G. and Sokoloff, K.L. (2002).
Exports and
Manufacturing Productivity in East Asia: A Comparative Analysis with
Firm-Level Data
. NBER Working Paper No. 8894.
Head, K. and Ries, J. (2003).
Heterogeneity and the FDI versus Export Decision of
Japanese Manufacturers.
Mimeo, University of British Columbia.
Heckman, J.J., Ichimura, H. and Todd, P. (1998). Matching as an Econometric
Evaluation Estimator.
Review of Economic Studies
65(2), 605-54.
Heckman, J.J., Lalonde, R.J. and Smith, J.A. (1999). The Economics and
Econometrics of Active Labor Market Programs. In: A. Ashenfelter and D.
Card (eds),
Handbook of Labor Economics
3, Amsterdam, 1866-2097.
Kraay, A. (1999). Exportations et Performances Economiques: Etude d’un Panel
d’Entreprises Chinoises
. Revue d’Economie du Développement
, 1-2/1999,
183-207.
Leuven, E. and Sianesi, B. (2003).
PSMATCH2: Stata Module to Perform Full
Mahalanobis and Propensity Score Matching, Common Support Graphing,
and Covariate Imbalance Testing
. Version 1.2.0,
http://ideas.repec.org/c/boc/bocode/s432001.html.
Marin, D. (1992). Is the Export-led Growth Hypothesis Valid for Industrialized
Countries?.
Review of Economics and Statistics
74(4), 678-88.
Marschak, J. and Andrews, W.H. (1944). Random Simultaneous Equations and the
Theory of Production.
Econometrica
12, 133-205.
Melitz, M. (2004). The Impact of Trade on Intra-Industry Reallocations and
Aggregate Industry Productivity.
Forthcoming in
Econometrica
.
Melitz, M. and Ottaviano, G.I.P. (2003).
Market Size, Trade, and Productivity
.
Working Paper, Harvard University.
Olley, S. and Pakes, A. (1996). The Dynamics of Productivity in the
Telecommunications Equipment Industry.
Econometrica
64, 1263-97.
Rosenbaum, P.R. and Rubin, D.B. (1983), The Central Role of the Propensity
Score in Observational Studies for Causal Effects.
Biometrika
70(1), 41-55.
Sachs, J. and Warner, A. (1995). Economic Reform and the Process of Global
Integration.
Brookings Papers on Economic Activity
1, 1-95, Washington
DC.
Wagner, J. (2002). The causal effects of exports on firm size and labor
productivity: First evidence from a matching approach.
Economics Letters
77(2), 287—92.
26
Appendix A. Estimation of Firm Productivities
Firm productivities are estimated assuming a Cobb-Douglas
production function with labour and capital as input factors. The
output measure used is firm value-added. The estimation equation
(in logarithmic form) is hence:
itititit ukly +⋅+⋅=
γβ
In this equation, the estimated error term uit proxies the logarithm of
plant-and time-specific total factor productivity. The problem usually
referred to as the simultaneity problem is that at least a part of the
TFP will be observed by the firm at a point in time early enough so
as to allow the firm to change the factor input decision. Profit
maximization then implies that the realization of the error term is
expected to influence the decision on factor inputs, rendering OLS
estimation inconsistent. In order to initialize the dynamic process
governing inputs and error terms, we have to assume the history
preceding the first observation in our sample as exogenous. Our
semiparametric estimation procedure following Olley and Pakes
(1996) involves two steps. In a first step, we assume that investment
and capital stock are linked by the equation
ititit IKK +−=
+)1(
1
δ
where K is capital stock and I is investment. Investment is then a
function of the capital stock and of the part
ϖ
it
of TFP that is
observed by the firm early enough to influence the investment
decision:
),( itittit kii
ϖ
=.
Defining the inverse function h( ) = i-1( ), we can write
ϖ
it
=h
t
(i
it
, k
it
)
and estimate
ititititit ekily ++⋅= ),(
φβ
where the function
φ(i
it
,k
it
) = γ k
it
+ h
t
(i
it
, k
it
)
is approximated by a
3rd order series estimator. The coefficient of logarithmic labour is now
consistently estimated.18 In a second step, we identify the capital
coefficient consistently by estimating the equation
itttititit ekgkly +⋅−+⋅=⋅− −− )( 11
γφγβ
18 Inverting the function i requires a monotonicity assumption regarding investment. In
contrast to other firm survey data, our investment data are very complete, making this
assumption seem reasonable in our case.
27
where g is an unknown function that is again approximated by a
third order polynomial expression in φt-1 and kt-1. The consistent
factor coefficient estimates allow us to construct the residuals of
equation (1). In this paper, productivity is used as a relative
measure, dividing the individual values over the mean of the
respective NACE2-industry and year.
28
Appendix B. Classification of Economic Activities
Our data contains firms in the manufacturing sector, as defined by
Nace-Classifications 15 to 36. This definition excludes natural-
resources-based activities such as agriculture, fishing, and mining,
utilities like the generation of electricity, water, recycling and the
construction sector. For our estimations, we divided the
manufacturing sector into 4 aggregate industries, as shown below.
NACE2-Classification of the Manufacturing Sector Industry
15 Manufacture of food products and beverages 1
16 Manufacture of tobacco products (no observations in our sample)
17 Manufacture of textiles 2
18 Manufacture of wearing apparel; dressing and dyeing of fur 2
19 Tanning and dressing of leather; manufacture of luggage, handbags,
saddlery, harness and footwear
2
20 Manufacture of wood and of products of wood and cork, except furniture;
articles of straw and plaiting materials
2
21 Manufacture of pulp, paper and paper products 2
22 Publishing, printing and reproduction of recorded media 2
23 Manufacture of coke, refined petroleum products and nuclear fuel 3
24 Manufacture of chemicals and chemical products 3
25 Manufacture of rubber and plastic products 3
26 Manufacture of other non-metallic mineral products 3
27 Manufacture of basic metals 3
28 Manufacture of fabricated metal products, except machinery and
equipment
3
29 Manufacture of machinery and equipment n.e.c. 4
30 Manufacture of office machinery and computers 4
31 Manufacture of electrical machinery and apparatus n.e.c. 4
32 Manufacture of radio, television and communication equipment and
apparatus
4
33 Manufacture of medical, precision and optical instruments, watches and
clocks
4
34 Manufacture of motor vehicles, trailers and semi-trailers 4
35 Manufacture of other transport equipment 4
36 Manufacture of furniture; manufacturing n.e.c. 4
