Content uploaded by Matthias Almus
Author content
All content in this area was uploaded by Matthias Almus on Apr 22, 2015
Content may be subject to copyright.
econstor
www.econstor.eu
Der Open-Access-Publikationsserver der ZBW – Leibniz-Informationszentrum Wirtschaft
The Open Access Publication Server of the ZBW – Leibniz Information Centre for Economics
Nutzungsbedingungen:
Die ZBW räumt Ihnen als Nutzerin/Nutzer das unentgeltliche,
räumlich unbeschränkte und zeitlich auf die Dauer des Schutzrechts
beschränkte einfache Recht ein, das ausgewählte Werk im Rahmen
der unter
→ http://www.econstor.eu/dspace/Nutzungsbedingungen
nachzulesenden vollständigen Nutzungsbedingungen zu
vervielfältigen, mit denen die Nutzerin/der Nutzer sich durch die
erste Nutzung einverstanden erklärt.
Terms of use:
The ZBW grants you, the user, the non-exclusive right to use
the selected work free of charge, territorially unrestricted and
within the time limit of the term of the property rights according
to the terms specified at
→ http://www.econstor.eu/dspace/Nutzungsbedingungen
By the first use of the selected work the user agrees and
declares to comply with these terms of use.
zbw
Leibniz-Informationszentrum Wirtschaft
Leibniz Information Centre for Economics
Almus, Matthias; Czarnitzki, Dirk
Working Paper
The effects of public R&D subsidies on firms'
innovation activities: the case of Eastern Germany
ZEW Discussion Papers, No. 01-10
Provided in Cooperation with:
ZEW - Zentrum für Europäische Wirtschaftsforschung / Center for
European Economic Research
Suggested Citation: Almus, Matthias; Czarnitzki, Dirk (2001) : The effects of public R&D
subsidies on firms' innovation activities: the case of Eastern Germany, ZEW Discussion Papers,
No. 01-10
This Version is available at:
http://hdl.handle.net/10419/24429
ZEW
Zentrum für Europäische
Wirtschaftsforschung GmbH
Centre for European
Economic Research
Discussion Paper No. 01-10
The Effects of Public R&D Subsidies
on Firms’ Innovation Activities:
The Case of Eastern Germany
Matthias Almus and Dirk Czarnitzki
Discussion Paper No. 01-10
The Effects of Public R&D Subsidies
on Firms’ Innovation Activities:
The Case of Eastern Germany
Matthias Almus and Dirk Czarnitzki
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/dp0110.pdf
The Effects of Public R&D Subsidies on Firms’
Innovation Activities: The Case of Eastern Germany
to appear in April 2003 in the Journal of Business and Economic Statistics
Matthias Almus and Dirk Czarnitzki
Centre for European Economic Research (ZEW), Mannheim
This version: August 2002
First version: February 2001
Abstract
This study analyzes the effects of public R&D policy schemes on
the innovation activities of firms lo c ated in Eastern Germany. The
main question in this context is whether public funds stimulate R&D
activities or simply crowd out privately financed R&D. Empirically,
we investigate the average causal effects of all public R&D schemes in
Eastern Germany using a non–parametric matching approach. Com-
pared to the c ase where no public financial means are provided, it
turns out that firms increase their innovation activities by about four
percentage points.
Keywords: Public Innovation Subsidies, Non–parametric Matching.
JEL Classification: C41, O31, O38.
address: Centre for European Economic Research (ZEW)
Department of Industrial Economics and International Management
P.O.Box 10 34 43, D–68034 Mannheim
phone: +49/621–1235–185, –158
fax: +49/621–1235–170
e–mail: almus@zew.de, czarnitzki@zew.de
1 INTRODUCTION
In 1998, the German federal government sp e nt about 2.2 billion Euro on pro-
moting R&D activities in the business sector. Given this large amount of
public R&D subsidies for innovation projects the question arises of whether
policy schemes stimulate private activities that produce positive externali-
ties, i.e. benefits to society.
The economic literature concerning external effects indicates that innova-
tion projects lead to market failures. Innovations are assumed to have pos-
itive external effects, but firms only launch privately profitable innovation
projects. Thus, there may be projects that would have p os itive benefits to
society, but do not cover the private cost. Therefore, these projects are not
carried out and the quantity of innovations is below the socially desirable
level. This circumstance is the main reason for governments to subsidize
private R&D projects. Public funding reduces the price for private investors
and thus the innovations are carried out. However, a firm always has an
incentive to apply for public R&D support, even if it could perform the
R&D projects using its own financial means. If public support is granted,
the firm then might simply substitute public for private investment. This
possible crowding–out effect between public grants and private investment
has to be taken into account when public authorities decide on the level of
their engagement in R&D support programs.
This study investigates the effects of R&D subsidies in Easte rn Germany
which is more than a decade after the break dow n of the Berlin Wall still
a transition economy. Public authorities have been trying to accelerate the
transition pro c es s from a planned economy to a market economy since the
German reunification in 1990. The efforts undertaken were and still are
enormous but many industrial firms are still struggling to survive. In most
regions of Eastern Germany, the number of producing firms per inhabitant
is far below the average of West Germany. Moreover, firms are mostly too
small, i.e. they have not reached the minimum efficient size of production
(MES). Furthermore, firms suffer from the collapse of “eastern markets”
which still induces severe sales difficulties. To overcome these difficulties
various programs and schemes have been set up.
1
For example, Ragnitz (2000) compares all subsidies granted in Eastern and
Western Germany. In relation to the labor force, the amount is twice as high
in Eastern Germany. Instead of the labor force, one can also consider sub-
sidies in relation to the gross domestic product. In this case, subsidies are
more than three times higher in Eastern Germany than in Western Germany.
According to calculations of Ebling et al. (1999), about 60% of innovating
firms in Eastern Germany received public R&D funding in 1996. This quote
is six times higher than in the western part of Germany. Moreover, there are
considerably more m eans spent on new firms to establish a certain amount
of small and medium sized firms (SME) that are important for a power-
ful market economy (Almus 2001). These figures make an examination of
public R&D schemes in Eastern Germany an interesting and necessary task.
Therefore, in this paper we analyze whether a crowding–out effect among
public R&D funds and privately financed R&D activities occurs in the East-
ern German economy.
2 THIS STUDY IN CONTEXT OF EXISTING
LITERATURE
Several empirical studies already exist on the effects of public R&D sub-
sidies. David et al. (2000) review the literature on the relation between
R&D subsidies and R&D expenditure on different levels of aggregation. All
studies reviewed aim to explore the sign and the magnitude of the “net”
effect of public policies. On industry or country level only 2 out of 14 em-
pirical studies report that public R&D funding crowds out private R&D
investment. The evidence is less clear at the firm level: 9 out of 19 studies
indicate substitutional effects, i.e. public funds crowd–out private invest-
ment either partially or even completely.
The difficulty of this kind of analysis are potential selection biases coming
from the public institutions that — depending on the applying firm and
the relevant R&D project — decide the recipients of the public funding
solely: “This makes public funding an endogenous variable, and its inclu-
sion in a linear regression will cause inconsistent estimates if it happens
to be correlated with the e rror term” (Busom, 2000: 114). Furthermore,
2
public institutions might support only those firms and R&D projects that
are expected to generate extensive economic spillover effects. To estimate
the “real” effects of public subsidies it is therefore necessary to address the
core evaluation question: How much would the subsidy receiving firms have
invested, if they had not participated in a public policy scheme? In fact,
only a few studies on the impact of R&D subsidies attempt to model this
counterfactual situation. Most of the studies surveyed in David et al. (2000)
do not pay attention to this kind of selection bias.
Recently, Wallsten (2000) considers a simultaneous equation model to pay
attention to the possible interdependence between public R&D funding and
R&D expenditure of firms. He investigates the Small Business Innovation
Research (SBIR) program and concludes that it is necessary to account for
possible endogeneity of federal R&D grants. According to the results of
the study SBIR awards crowd out firm–financed R&D spending dollar for
dollar (full crowding–out). The subsidies do neither have an effect on R&D
activities nor on employment. However, he mentions another possible and
important impact of public funding: “[...] while the grants did not allow
firms to increase R&D activity, they instead allowed firms to continue their
R&D at a constant level rather than cutting back.” (Wallsten, 2000, p. 98).
Busom (2000) explores the problem of selection bias by applying a two stage
econometric treatment model where the first stage consists of estimating a
probit model on the participation probability in public funding programs.
In the second stage, the R&D activity is regressed on several covariates
including a selection term which accounts for the different prope nsities of
firms to be publicly funded. This second equation is estimated separately
for participants and non–participants. The difference in expected R&D ex-
penditure of both groups is according to this approach the result of public
funding. Busom concludes that for the majority of firms in her sample public
funding induced more R&D activities, but for 30% of participants complete
crowding–out effects cannot be ruled out.
Lach (2000) investigates the effects of R&D subsidies granted by the Israeli
Ministry of Industry and Trade on local manufacturing firms. He applies
different estimators, such as the before–after–estimator, the difference–in–
3
difference estimator and different dynamic panel data models. Although
Lach finds heterogenous results from different models applied, he finally
concludes that subsidies do not crowd out company financed R&D expendi-
ture completely. Their long–run elasticity with respect to R&D subsidies is
0.22.
Other microeconomic approaches do not focus on crowding–out effects but
take different output measures into consideration: for example, the effects
of subsidies on patent applications, productivity, fixed asset investments,
returns on capital, returns on sales and growth of sales or employment (see
Klette 2000 for a comprehensive survey).
This study focuses on the crowding–out issue and introduces another empir-
ical tool to the literature on examining the effects of public R&D funding.
We apply a non–parametric matching approach that goes back to the model
of p otential outcomes developed by Roy (1951) and Rubin (1974). These
matching approaches were extensively applied in the literature on the eval-
uation of labor market policies, e.g. the evaluation of active labor market
programs (ALMP) or qualification measures (LaLonde 1986, Dehejia and
Wahba 1999, Lechner 1999, Heckman et al. 1999). In these cases, people
are the subject of the examination and research questions include whether
wages, salaries or the probability of being hired or re–employed increase
if people take part in a specific measure or program. The non–parametric
matching approach applied here can clearly identify the effect that goes back
to the receipt of public R&D funding, since we are able to approximate a
situation where no differences exist between subsidized and non–subsidized
firms with respect to characteristics that influence the probability to receive
public support and to carry out private R&D. According to Hausman (2001)
the matching metho dology leads to more robust estimates of the treatment
or causal effect compared to alternative approaches.
A major advantage of this study is the ability to identify exactly whether a
firm received any subsidies for innovative projects. All programs launched by
public authorities are incorporated and so the approach applied can reflect
the effects of public R&D p olicy schemes collectively and is not restricted to
a particular measure. Many other studies only deal with one specific public
4
R&D program and cannot control for possible effects of other publicly funded
research. In contrast, we can distinguish recipients and non–supported firms
in the sample exactly. Our control group contains only firms which did not
receive any public R&D grants. This is not the case for several other stud-
ies that analyze one specific R&D program but are not able to control for
other sources of public funding. However, this advantage has its price: We
are not able to track in which particular program a firm participated. We
only observe whether a firm participated in any public R&D scheme under
consideration. Therefore, we do not describe the R&D programs in more
detail. Of course, the treatments were targeting different types of firms or
aims and thus heterogeneous treatments exist. Hence, our study can only
be seen as broad evidence on the ove rall R&D p olicy in Eastern Germany
and is only able to discover average effects over different s chemes.
3 DATA
The data used is taken from the Mannheim Innovation Panel (MIP) con-
ducted by the Centre of European Economic Research (ZEW) on behalf of
the German Federal Ministry for Education and Research (cf. Janz et al.
2001 for a more detailed description of the MIP database). The MIP is a
German survey on innovation activities in the business sector. It formed the
German part of the Community Innovation Survey (CIS) of the European
Commission in 1993, 1997 and 2001. Since 1993 information from about
2,500 German manufacturing firms has been collec ted in the MIP annually.
We use data from the surveys in 1995, 1997 and 1999, i.e. the informa-
tion collected corresponds to the firms’ activities in 1994, 1996 and 1998.
Firms in the survey are from almost the whole business sector and can be
classified according to the European standard classification NACE. We use
the manufacturing sector and, thus, firms in the sample belong to twelve
industries that are characterized by dummies in the empirical analysis (see
Table 5 in the appendix). Note that only firms with at least five employees
are sampled in the MIP. In critique of former studies, Lichtenberg (1984)
argues that the results of evaluations are often biased because the data used
is mainly comprised of observations on large firms. The MIP data overcome
this problem as there are many observations on small and medium sized
5
firms (see des criptive statistics). The sample contains 925 observations on
innovating firms located in Eastern Germany from which 622 participated
in public R&D schemes. Note that we use our database not as panel data
but as three cross–sections. Most firms, i.e. more than 70%, are only ob-
served once in the sample. Only about 8% are included in every of the three
cross–sections. Table 1 contains descriptive information on participating
firms and the potential control group of non–participants. According to the
Oslo–Manual guidelines (Eurostat and OECD 1997) innovators are defined
as firms which have introduced at least one product or process innovation
in recent three years. A product or process innovation is defined as follows:
“Technological product and process (TPP) innovations com-
prise implemented technologically new products and processes
and significant technological improvements in products and pro-
cesses. A TPP innovation has been implemented if it has been
introduced on the market (product innovation) or used within
a production process (process innovation). TPP innovations in-
volve a series of scientific, technological, organizational, financial
and commercial activities. The TPP innovating firm is one that
has implemented technologically improved products or processes
during the period under review.” (Eurostat and OECD 1997, p.
47)
As potential outcome variable in the empirical analysis, the R&D intensity
is considered, i.e. the ratio of R&D expenditures to sales (multiplied by
100). We separate our sample with respect to the participation in public
R&D schemes into the treatment group, i.e. subsidized firms, and potential
control group. The empirical analysis then tries to assess whether firms that
received public R&D funds in 1994, 1996 or 1998 have on average a higher
R&D intensity compared to firms that did not receive public means in the
period under consideration. There are three time periods under evaluation
and a firm may belong to the group of subsidized firms (treatment group)
in one, two or all three periods. However, we only allow firms to enter the
potential control group if they have previously not participated in any of the
R&D support programs. Hence, all firms that received public R&D funds in
1994 but not in subsequent years under examination or in 1996 but not in
1998 are excluded from the potential control group to avoid biased results.
An important part of the empirical analysis is to estimate the probability of
6
a firm to receive public funds given a number of observable characteristics
which also have an influence on the succes s variable, i.e. the R&D intensity.
Therefore, several control variables that are used in the empirical analysis
are presented briefly in the following paragraphs.
The log of the number of employees and its square take account of possible
size effects. A p ote ntial concern of using the number of employees is the
fact that firms which receive subsidies may hire R&D staff and thus their
employment increases. This would cause some endogeneity among the re-
ceipt of public funding and firm size. Therefore, it would b e preferable to
use the lagged number of employees of the year prior to participation in
public policy schemes but we do not have the required information in our
database. Howeve r, we think that the possible endogeneity problem is not
severe in our study for two reasons: First, there are only a few programs
elaborated towards increasing the R&D staff directly. Second, R&D staff as
a proportion of all employees of the firm amounts to less 5% on average for
the firms in the database, where this figure is quite stable over time. Hence,
R&D subsidies may influence the number of R&D staff in some cases but
this change is small compared to the number of all employees. These two
arguments weaken the concern of a potential endogeneity between the re-
ceipt of R&D subsidies and the number of employees.
Eleven industry dummies control for cross–sectional differences, e.g. differ-
ent technological potential in various industries. Two cohort dummies shift
inter–temporal effects. Another important factor that might have an influ-
ence on the probability of funding as well as on the success measure is market
competition. T hus, several variables control for competitive impacts: the
market share variable measures the firms’ sales in relation to the industries’
sales measured on the NACE three digit level. The import ratio measured
on the two digit sectoral level captures the competitive pressure of foreign
firms on the market. Moreover, we consider the firms’ export related sales
divided by total sales to measure foreign competition. The sellers concen-
tration on the domestic market is also take n into account. This is measured
as the concentration ratio CR6, i.e. the sum of market shares of the indus-
tries’ six largest firms. Capital intensity, i.e. the ratio of tangible assets
per employee, is included in the analysis to c ontrol for different technologies
7
used in the production process. Moreover, we incorporate the firms’ age.
It is often claimed that older firms are more reluctant to pursue innovation
and, thus, one may argue they are less likely to apply for public research
programs. The foundation of a firm usually induces innovation activities
and, hence, young firms are expected to be more lively regarding R&D.
The legal form indicates the attitude of the firm (owner) towards risk and
also the chance to enter public R&D programs. Hence, the dummy vari-
able ‘legal form’ separates the sample in firms with liability limiting legal
forms (joint stock company [AG], non–public limited liability firm [GmbH]
or commercial partnership with a non–public limited liability firm [GmbH
& Co.KG]). For these firms the legal form dummy is zero. Using these legal
forms owners can minimize their risk up to a certain amount and thus have
higher incentives to pursue more risky projects (Stiglitz and Weiss 1981).
The dummy is one for firms with remaining legal forms (joint partnerships
etc). Companies with limited liability have much better options to receive
public subsidies because if firms apply for public grants, they have to prove
that they maintain an operating industrial plant. Firms with a liability lim-
iting legal form have to be recorded in the trade register in Germany which
means a publicly available information exists that this firm is already doing
its business. Companies with other legal forms have to prove this within
their application for public grants and the ministry official has to inspect
this on her or his own. Due to the fact, that ministy officials may behave
risk averse, companies with limited liability are possibly favored because
they have already proved their credibility.
To control for technological prowess or previous R&D experience a dummy
variable, indicating if firms have R&D departments, enters the analysis.
The inclusion of this dummy holds the potential of creating an endogeneity
problem. However, this would only be the case if firms in the sample were
establishing new R&D departments as a result of the receipt of public sub-
sidies. As there are no public R&D schemes in Germany which explicitly
support the founding of whole R&D departments, the endogeneity problem
is unlikely to occur. However, the R&D department dummy reflects the ab-
sorptive capacity and R&D experience of firms. The use of other variables
is not possible with our data: Unfortunately, using the (share of) R&D
8
personnel would cause endogeneity problems because there are some policy
schemes which promote hiring R&D staff. Other indicators of absorptive
capacity like lagged values of R&D expenditure are not available.
Finally, we incorporate dummy variables that indicate if the observed firm is
a subsidiary of a foreign or West German firm. This is done for two reasons:
There are many policy schemes especially for small and medium sized firms
(SME). However, if a firm is an SME but also belongs to a group with a
large parent company, this firm would not be accepted to participate in
policy schemes designed for SME. Moreover, many schemes are exclusively
for Eastern German firms, but if the parent company is a western one, the
subsidiaries are not allowed to enter in programs for Eastern German firms.
Hence, the dummy variables ‘Western German parent company’ and ‘foreign
parent company’ should capture these effects.
4 IDENTIFICATION AND MATCHING
4.1 Causal Effects and Potential Results
The situation to be examined is typical for an evaluation. All firms in the
database can be separated with respect to the receipt of public R&D subsi-
dies. This leads to a non–experimental setting since the receipt of subsidies
is not random. There are se veral differences between the groups of firms
with and without R&D subsidies as the upcoming empirical analysis will
reveal. The receipt of public R&D s ubsidies finally leads to a pote ntial out-
come Y
1
for the firms that received subsidies and Y
0
for the non–recipients.
The approach that is used to measure the difference between groups, i.e. the
causal effect, goes back to the model of potential outcomes by Roy (1951)
and Rubin (1974). Rubin defined the term causal effect as: “[. . .] the differ-
ence between the likely outcome of a person’s participation in the measure
and the likely outcome of a person’s non participation.” The participation
of firm i in any R&D scheme is denoted with S
i
= 1 and S
i
= 0 otherwise.
The evaluation aims to calculate the causal effect of public R&D schemes in
the subsidised firms’ view, i.e. the study concentrates on the causal effect
θ
1
that results from receiving R&D subsidies:
θ
1
:= E[Y
1
− Y
0
|S = 1] = E[Y
1
|S = 1] − E[Y
0
|S = 1] (1)
9
where E[•] in equation (1) represents the expectation operator. The causal
effect then indicates w hether public R&D support has a positive impact
on the private R&D intensity. However, the outcome E[Y
0
|S = 1] is by
definition not observable, since non–subsidised firms cannot be observed in
the case of R&D subsidy receipt. The first outcome E[Y
1
|S = 1] can be
estimated unbiased as the mean value of the outcome variable representing
firms that received subsidies. To identify E[Y
0
|S = 1] we have to incorporate
further assumptions.
4.2 Identification
E[Y
0
|S = 1] cannot simply be calculated as arithmetic mean of the non–
recipients since:
E[Y
0
|S = 1] 6= E[Y
0
|S = 0] . (2)
This condition would only be valid in the case of an experiment where partic-
ipants and non–participants are randomly assigned to the measure. The de-
scriptive analysis, however, shows that subsidized and non–subsidized firms
in our sample differ in various important characteristics. Due to selection
processes on the part of the authorities that decide how to distribute the
funds among applicants, the group of firms that received ass istance is a
special and selective one. Moreover, firms have different information and
different access to information regarding possibilities of application for pub-
lic funds. This may be a further source of potential selection.
Rubin (1977) introduces the conditional independence assumption (CIA) to
solve the problem arising in equation (2). This condition means that par-
ticipation (receipt of subsidies) and potential outcome (R&D intensity) are
independent for individuals with the same set of exogenous characteristics
(X = x
i
):
(Y
0
, Y
1
)⊥ S|X = x (CIA) . (3)
The condition helps to overcome the problem that E[Y
0
|S = 1] is unob-
servable. If CIA is valid, E[Y
0
|S = 0, X = x
i
] can b e used as a measure of
potential outcome for the R&D recipients (Lechner 1998). CIA, however,
is only plausible if all variables that influence the outcome Y
0
or Y
1
and
10
the participation status S are known and available in the data set. While it
is not possible to test the validity of CIA formally (see Almus et al. 1999),
the MIP contains a rich set of information that we believe makes the CIA a
reasonable approximation. If CIA is correct the equation:
E[Y
0
|S = 1, X = x] = E[Y
0
|S = 0, X = x] (4)
holds, which means that the outcome of non–participants can be used to
calculate the average outcome for the participants in an unbiased way pro-
vided that there are no systematic differences between firms with and with-
out public R&D subsidies. Then, the causal effect of public subsidization in
equation (1) changes to:
θ
1
:= E[Y
1
|S = 1, X = x] − E[Y
0
|S = 0, X = x] (5)
which can be estimated unbiased using the means of both groups (Lechner
1998). The next step requires a search for pairs of non–subsidized and sub-
sidized firms that do not differ in characteristics contained in the vector X.
Here, the study deviates from other examinations. Normally, there are more
firms or individuals in the potential control group compared to the group
of treated individuals or firms. Our database, however, has about twice
as many firms that received public R&D subsidies than non–recipients due
to the special situation in Eastern Germany after reunification. Then, the
matching approach assigns to every subsidized firm a similar non–subsidized
counterpart. Using this approach we do not waste information of subsidized
firms. However, a non–subsidized firm may be matched to more than one
recipient of R&D subsidies.
4.3 Non–parametric Matching
Rosenbaum and Rubin (1983) point to the fact that a large number of ex-
ogenous characteristics is required to ensure the validity of the CIA. The
vector x
i
containing the exogenous variables of firm i therefore has a high
dimension. This impedes the estimation of the causal effect, since it is
almost impossible to find subsidized and non–subsidized firms that have
exactly the same values in the exogenous variables if there are many to con-
sider. Fortunately, the vector of exogenous variables x
i
can be condensed
into a single scalar measure to solve this problem, the so called propensity
11
score. This measure represents the probability that a given firm i has re-
ceived public R&D subsidies at all given a set x
i
of individual characteristics
P r(S
i
= 1|X = x
i
). Rosenbaum and Rubin (1983) show that if the CIA is
fulfilled, it is sufficient to condition on the propensity score to ensure statis-
tical independence between potential outcome and receipt of R&D subsidies.
There are several forms of conditioning that can be summarized under the
term balancing scores (Rosenbaum and Rubin 1983, Lechner 1998). Balanc-
ing scores cover a wide range of measures starting from the most complex
X = x
i
to the propensity score Pr(S
i
= 1|X = x
i
) as most simple form.
This analysis uses the unbounded propensity score x
0
i
ˆ
β as single matching
criterion.
Beside the independence between potential outcome (firm specific R&D in-
tensity) and participation status (receipt of public R&D funds) the identifi-
cation of the causal effect depends on a further condition. Individual causal
effects may not be influenced by the participation status of other firms, i.e.
the absence of indirect effects (SUTVA [stable unit treatment value assump-
tion] condition) (Angrist et al. 1996). SUTVA constitutes a potential caveat
of the analysis but since all R&D programs in Eastern Germany are consid-
ered these possible indirect effects should not cause biased results: The firms
compete for the means on many sub–markets (various schemes). Regarding
a possible demand shift for R&D inputs and thus a change in factor prices,
we do not believe that public policy schemes have a remarkable effect. In
our opinion, the market for R&D inputs can be seen as a national market
rather than several regional ones. Admittedly, a proportion of 60% of inno-
vating firms was subsidized in Eastern Germany but when looking at whole
Germany this proportion is rather small, because less than 14% of German
innovators are located in Eastern Germany. In Western Germany only about
15% of innovating firms receive any public funding. Thus, the majority of
German innovators in the manufacturing sector does not participate in pub-
lic R&D schemes. Moreover, the amount of subsidies for the recipients is low
compared to their private investments. For example, in 1999, firms spent
on R&D activities about DM 60 billion in Germany, while the public R&D
subsidies of the federal government amounted to about DM 2 billion for
civilian R&D (BMBF 2000). Unfortunately, there are no figures available
for Eastern Germany only. However, as the share of subsidies is only about
12
3%, it seems to be unlikely that public R&D schemes have a significant influ-
ence on prices for R&D inputs. He nce, the SUTVA is assumed to be fulfilled.
Other approaches that can be used to estimate the causal effect in non–
experimental settings exist. The most often applied are (for a comprehensive
overview cf. Heckman et al. 1999):
- The “difference–in–differences” method (Ashenfelter 1978, Ashenfelter
und Card 1985) became popular with the availability of panel data.
Here, potential selection biases stemming from observable time invari-
ant variables vanish in the linear model if differences are calculated
over time (Fitzenberger und Prey 1998).
- Complete econometric selection models simultaneously estimate par-
ticipation and success of the program or measure. These models de-
pend on restrictive assumptions regarding the error terms and their
distribution that often cannot be interpreted economically. Therefore,
these models have often been criticized (Ashenfelter und Card 1985).
However, Heckman und Hotz (1989) point out that the application of
parametric models leads to satisfying results.
- Parametric instrument variable estimators which have increasingly
gained attention in recent years may be seen as variant of parametric
selection models (Angrist et al. 1996).
All these approaches have their advantages and disadvantages and there
are currently no guidelines when to use statistical matching or econometric
evaluation models. “[...] Thus the choice of an appropriate econometric
model critically depends on the data on which it is applied” (Heckman et
al. 1996). Moreover, Heckman und Hotz (1989) c onclude that “[...] there is
no objective way to choos e among alternative nonexperimental estimators.”
We finally apply a matching approach since the data set has comprehensive
information on the firms, thus enabling us to find a “perfect twin”, i.e. a sim-
ilar control observation for every s ubsidized firm in the upcoming matching
process. Moreover, Hausman (2001) states that matching approaches lead
to more robust estimates of the treatment effect compared to other methods.
13
5 EMPIRICAL ANALYSIS
5.1 Initial Situation and Probit Estimation
5.1.1 Pre–match Situation
The data set contains 625 firms (N
1
) that received public R&D subsidies.
Moreover, there are 303 firms (N
0
) that did not receive any public R&D
subsidies. Table 1 shows that there are significant differences in the means
of several characteristics between both groups (see columns 2 and 3). This
indicates that the group of firms that received public R&D subsidies is a
selective one. The decision of the firm to apply for public assistance as well
as the selection mechanisms on the part of the authorities which distribute
the means generate a group of firms with special characteristics. Therefore,
a comparison of the firm specific R&D intensities using the initial data set
would lead to biased results due to the differences between both groups.
5.1.2 Specification Tests and Probit Estimation
The best and easiest way to find a counterpart for every firm that received
public R&D subsidies is to select the non–subsidized one with exactly the
same values in the selected matching variables (see Table 1), i.e. a perfect
twin. But the relatively large number of these variables and the availability
of only about 300 firms in the p otential control group impedes this approach.
Matching methods which recently became popular in labor market evalua-
tion studies represent a powerful alternative avoiding these difficulties (Lech-
ner 1998). Rosenbaum and Rubin (1983) point out that matching “[. . .] is
a method for selecting units from a large reservoir of potential comparisons
to produce a comparison group of modest size in which the distribution of
covariates is similar to the distribution in the treated group.”
The matching algorithm used corresponds closely to the one applied by
Lechner (1998). To reduce the multidimensional problem arising from the
relatively large number of covariates to a one–dimensional, initially a pro-
bit model is es timated. The decision whether the firm has received public
assistance (S
i
= 1) or not (S
i
= 0) serves as the endogenous variable:
E[S
i
|X = x
i
] = P r(S
i
= 1|X = x
i
) = Φ(x
0
i
β) ∀ i = 1, . . . , N
0
+ N
1
(6)
14
Table 1: Mean Comparisons of subsidised firms, firms from the potential
control group without subsidisation and the selected control groups
subsidized non-subsidized
a
selected control
b
firms firms firms
industry 1 0.053 0.142
∗
0.053
industry 2 0.061 0.056 0.061
industry 3 0.023 0.066
∗
0.023
industry 4 0.084 0.066 0.084
industry 5 0.068 0.089 0.068
industry 6 0.058 0.086 0.058
industry 7 0.143 0.214
∗
0.143
industry 8 0.214 0.073
∗
0.214
industry 9 0.122 0.046
∗
0.122
industry 10 0.093 0.059 0.093
industry 11 0.042 0.056 0.042
industry 12 0.040 0.046 0.040
cohort dummy 1994 0.320 0.558
∗
0.342
cohort dummy 1996 0.350 0.300 0.341
cohort dummy 1998 0.330 0.142
∗
0.317
number of employees 191.8 136.3 178.6
export ratio 0.171 0.099
∗
0.174
market share 0.385 0.278 0.350
import ratio 0.209 0.180
∗
0.209
sellers concentration (CR6) 0.185 0.169 0.186
West German parent company 0.196 0.191 0.220
foreign parent company 0.048 0.076 0.056
firm age 5.963 7.376 6.564
capital intensity 0.095 0.104 0.097
legal form 0.058 0.092 0.069
R&D department 0.603 0.248
∗
0.592
propensity sc ore 0.817 0.044
∗
0.801
observations 622 303 622
different control obse rvations / / 157
Notes:
∗
indicates that the means differ with statistical significance in a two–tailed
t–test at the 5% level between the supported firms (column 2) and either
firms from the potential control group (column 3) or from the selected control
group (columns 4).
a
Non–subsidised firms of the initial sample, i.e. prior to the matching procedure.
b
Selected non–subsidised firms, i.e. based on the matching procedure.
The vector x
i
contains the set of characteristics that potentially influence
the probability of receiving public R&D subsidies. These have been intro-
15
duced in section 3. Φ(•) is the cumulative density function of the standard
normal and β is the parameter vector to be estimated. N
1
and N
0
define
the number of assisted and non–assisted firms, respectively.
Tests on normality and heteroscedasticity have been carried out to find
potential misspecifications since these would lead to inconsistent probit es-
timates. We use Lagrange multiplier (LM) tests to check if misspecifications
of the distributional assumptions (non–normality, heteroscedasticity) exist
(cf. Verbeek 2000). The results of the heteroscedasticity tests are given in
Table 2. The statistics are χ
2
distributed with as many degrees of freedom
as variables to be tested for heteroscedasticity. The tests do not reject the
null hypothesis that error terms are homoscedastic at the 5% level of sig-
nificance. Moreover, the normality assumption cannot be rejected at the
5% level of significance in a χ
2
test with 2 degrees of freedom (see in Table
2). This test examines whether skewness and kurtosis are characteristic of
a normal distribution.
Table 2: He terosc edasticity and normality tests
variable degrees of freedom statistic prob–value
industry dummies 11 13.290 0.275
cohort dummies 2 4.245 0.120
size groups 5 5.050 0.410
export ratio 1 0.900 0.343
market share 1 0.015 0.903
import ratio 1 0.177 0.674
sellers concentration 1 0.070 0.791
parent company 2 1.096 0.578
1/age 1 0.471 0.492
capital intensity 1 0.213 0.644
legal form 1 0.000 0.984
R&D department 1 2.005 0.157
normality 2 4.941 0.085
number of observations 925
Hence, no indication of potential misspecification of the homoscedastic pro-
bit model are found. Thus, the results can be used for making inferences
and the upcoming matching process. Table 3 contains the estimated param-
16
eters which will be interpreted briefly at first. In addition to the estimated
parameters the table contains the marginal effects which are normally used
to interpret the results. Here, the effect of marginal changes of an exoge-
nous variable on the probability to receive subsidies c an b e examined. The
marginal effects for the probit model are calculated according to Greene
(2000) in the following way:
∂E[S| X = x]
∂x
k
=
∂P r(S = 1| X = x)
∂x
k
=
∂Φ(x
0
β)
∂x
k
= φ(x
0
β)β
k
. (7)
In equation (7) φ(•) is the probability density function of the standard nor-
mal.
In the probit estimation, several industry dummies, the cohort dummies,
the firm size as well as the fact that the potential parent company is located
abroad have a significant influence on the probability to receive public R&D
subsidies. Moreover, the sellers concentration as well as the existence of an
R&D department significantly determine the probability of b eing subsidized.
No a priori considerations were made regarding the influence of the industry
dummies. But it turns out that industries that are rather technology inten-
sive (industries 4, 8 to 11) have ceteris paribus a higher probability to receive
subsidies. The cohort dummies indicate that in subsidization periods 1996
and 1998 firms had a higher probability to receive subsidies compared to
the reference period 1994. The effects amount other things equal to about
20 and 25 percentage p oints. The existence of a foreign parent company is
connected with a decrease of the probability to receive public R&D subsi-
dies c.p. by about 26 percentage p oints. This indicates that German firms
without foreign links are the main focus of public support. The insignificant
influence of a West German parent company further supports this finding.
Firm size is a further determinant that significantly influences the subsidisa-
tion probability. The larger the firm the better its chances to receive public
funds. This is mainly due to information advantages, better capacities to
carry out R&D as well as the existence of more staff and capacity to apply
for the funds. According to the marginal effects an increase of the firm size
by 10% would raise the probability to receive subsidies by about 2.2 percent-
age points. An existing R&D department has not surprisingly a significant
positive effect on the probability to receive subsidies. The existence raises
17
Table 3: Re sults of the probit estimation
variable coefficient t-value marg. eff.
a)
t-value
industry 2 0.242 0.620 0.077 0.660
industry 3 0.242 0.860 0.077 0.930
industry 4 1.015 3.110
∗
0.247 4.940
∗
industry 5 0.450 1.920 0.135 2.240
∗
industry 6 0.324 1.390 0.101 1.540
industry 7 0.355 1.850 0.112 2.010
∗
industry 8 0.880 3.920
∗
0.241 5.080
∗
industry 9 1.582 4.010
∗
0.318 8.480
∗
industry 10 0.723 2.850
∗
0.197 3.760
∗
industry 11 1.095 2.330
∗
0.250 4.280
∗
industry 12 0.038 0.130 0.013 0.130
cohort 1996 0.621 5.420
∗
0.196 5.880
∗
cohort 1998 0.850 6.190
∗
0.251 7.420
∗
ln(employees) 0.641 2.580
∗
0.218 2.580
∗
ln(employees)
2
-0.051 -1.830 -0.017 -1.820
capital intensity -0.196 -0.450 -0.067 -0.450
1/age 0.618 1.170 0.210 1.170
West German parent company -0.223 -1.680 -0.079 -1.630
foreign parent company -0.680 -3.280
∗
-0.258 -3.160
∗
export ratio 0.004 1.550 0.001 1.550
import ratio 0.011 1.160 0.004 1.160
sellers concentration -0.022 -2.970
∗
-0.008 -2.970
∗
market share -0.006 -0.210 -0.002 -0.210
R&D department 0.681 6.300
∗
0.228 6.580
∗
legal form 0.122 0.640 0.040 0.660
intercept -2.459 -4.140
∗
/ /
Pseudo R
2
0.202
observations 925
Notes:
∗
indicates statistical significance at the 5% level.
a)
∂S/∂x is for dummy variables the discrete change from 0 to 1.
The marginal effects will be calculated at the means of the variables.
the probability by about 23 percentage points. Finally, the legal form does
not influence the probability to receive subsidies.
After estimation of equation (6) the unbounded propensity score x
0
i
ˆ
β is cal-
culated for every observation. This measure is used in the procedure to find
the counterparts for every subsidized firm. We prefer the unbounded rather
18
Note: Scores (x
0
i
ˆ
β) based on the probit model.
Figure 1: Frequency distribution of the unbounded propensity scores of the
initial data set
than the bounded propensity scores Φ(x
0
i
ˆ
β) because it has preferable dis-
tribution properties (Hujer et al. 1997). We also used Φ(x
0
i
ˆ
β) as matching
criterion but there were only marginal changes in the results of the following
matching process. Figure 1 shows frequency distributions of the unbounded
propensity scores x
0
i
ˆ
β of both firm groups for the initial data set. They
fulfill an important assumption for the matching process, since both graphs
overlap to a great extent, hence indicating similar distributions of the two
groups (Lechner 1998).
5.2 Non–parametric Matching
The general matching process applied proceeds as follows :
1. Separate the observations with respect to their status of public R&D
subsidy receipt.
2. Select a firm i that received public R&D funds.
3. Take the unbounded propensity score x
0
ˆ
β. In many empirical studies
one wants to balance the participiants and control observations with
regard to more characteristics than the propensity score. Firm size is
an example. Therefore one uses, additionally to the propensity score,
19
a vector ν (where ν is a subset of x) that contains important matching
variables. This variant is called hybrid matching (cf. Lechner 1998).
4. Then one calculates a proper measure of metric distance, e.g. the
Mahalanobis distance. Let:
d
ij
= (x
0
i
ˆ
β, ν
i
)
0
− (x
0
j
ˆ
β, ν
j
)
0
∀ j = 1, . . . , N
0
for every combination of the R&D recipient i and every firm from the
potential c ontrol group j. Then calculate the Mahalanobis distance:
MD
ij
= d
ij
0
Cov
−1
d
ij
∀ j = 1, . . . , N
0
to find the nearest neighbour. Cov represents the covariance matrix
based on the controls, i.e. firms that did not receive public subsidies.
5. After calculating the distance, one possibly wants to impos e some
restrictions on the neighborhood:
- A required criterium to be a neighbor of participant i may be
that a p otential control firm is recorded in the same industry
classification.
- One shortcoming of the nearest neighbor matching so far is that
always a neighbor is picked, even if the metric distance to the i–th
control observation is very large. To prevent too large distances,
it is possible to define a confidence interval of the propensity
score and other matching variables on basis of the participant
group in which a potential control observation should be included.
This is called calipre matching and was indroduced by Cochran
and Rubin (1973). Hujer et al. (1997) give an example for this
method.
6. The firm j from the potential control group with the smallest Maha-
lanobis distance serves as c ontrol observation in the following success
analysis. The comparison observation is drawn randomly if more than
one firm attains the minimum Mahalanobis distance. If no potential
control observation remains in the pool after applying the restrictions
described in the previous step, firm i is bypassed and no match can be
made.
20
7. Remove the i–th firm from the pool of firms that received subsidies but
return the selected control observation to the pool of control observa-
tions. This is done because of the relatively small number of control
firms. Using different data, i.e. a large potential control group, one
could also draw without replacement. In this case, it would be impor-
tant to draw the participants one after the other randomly from the
treatment group.
8. Repeat steps 2 up to 7 to find matched pairs for all recipients.
Following matching technique is applied in this paper: We only use the
propensity score and impose the restriction that potential controls have to
be recorded in the same industry classifications as the participants. If the
matching results are not satisfactory, one would proceed with additional
variables in the matching function. However, it turned out that using the
unbounded propensity sc ore as only matching criterion is already sufficient.
Table 1 measures the statistical “similarity” of the observations that remain
after the matching procedure. Column 2 contains the means of the variables
of the firms with R&D subsidies and c olumns 4 the means of the assigned
firms without such subsidies. Matching is regarded as successful if the means
of the relevant variables in both groups do not differ significantly. Note that
we found for every participant a neighbor within the confidence interval
defined by the calipre restriction with regard to the propensity score. As
indicated by a t–test, the differences of the means are small and not statis-
tically significant at the 5% level for all variables. Moreover, the unbounded
propensity score x
0
i
ˆ
β as a summary measure of various variables does not
significantly differ between both groups, indicating a good fit of the match-
ing algorithm applied. 157 out of the 303 potential control observations are
used for the selected control group. This means that each se lecte d control
group observation is on average assigned to four subsidized firms.
Figure 2 contains kernel density estimates of the unbounded propensity
scores x
0
i
ˆ
β for both groups. The Epanechnikov kernel density estimates
instead of histograms serve as tool to show the similarity in the relative fre-
quencies (probability density) since both groups contain the same number
of observations after the matching proce ss (c.f. Silverman 1986). The are
nearly no differences on the left and the middle part of the distribution. Due
to the small number of non–subsidized firms on the right tail (see Figure 1)
21
Note: Scores (x
0
i
ˆ
β) based on the probit model.
Figure 2: Density distribution of the unbounded propensity scores after the
matching process
it is difficult to find adequate matches pairs. All in all, the figure underlines
the quality of the matching procedure.
6 CAUSAL EFFECTS
The success of public R&D subsidies is evaluated by comparing the average
firm specific R&D intensities between the groups of subsidized and non–
subsidized firms, i.e. Y
1
i
and Y
0
i
. The unbiased estimator for the causal
effect
ˆ
θ
1
is the difference of the means between both groups
ˆ
θ
1
=
1
N
1
N
1
X
i=1
Y
1
i
−
N
1
X
i=1
Y
0
i
. (8)
R&D subsidy programs have on average a positive impact on the firm spe-
cific R&D intensity if the causal effect
ˆ
θ
1
is significantly greater than zero.
The programs do not generate positive effects if
ˆ
θ
1
is statistically insignifi-
cant. Finally, subsidized firms perform worse than firms without subsidies
if the causal effect is significantly smaller than zero. This means that non–
subsidized firms undertake on average more R&D efforts (measured with the
R&D intensity) than firms that received funding within the programs under
22
evaluation.
The test on the effect is usually carried out by means of a simple t–statistic.
In this case, however, the ordinary t–value is biased upwards because it does
not take into account that the mean of the outcome variable of the control
group is not a result of a random sampling but an estimation: it is based on
the estimated propensity s cores and the non-parametric matching procedure.
Thus, the usual t–statistic may be misleading for making inferences. To
remove the bias of the t–statistic, the method of bootstrapping is applied,
i.e. we simulate the distribution of the mean outcome of the control group
by repeated sampling (for a sketch of bootstrapping, see Greene 2000 or
Efron and Tibshirani 1993 for a comprehensive discussion):
- A random sample with replacement is drawn from the original sample,
which has the same size as the original one.
- Afterwards, we estimate the probit model again and perform a new
matching with this sample and record the mean difference
ˆ
θ
1
after the
procedure.
- The whole process is repeated 200 times.
- Subsequently, we receive a simulated distribution of mean differences
between the participants and their controls. This empirical distribu-
tion can subsequently be used to calculate a standard error and, thus,
an unbiased t–statistic.
Applying equation (8) leads to an average R&D intensity of about 6.6 (2.6)
% for the subsidized (non–subsidized) firms. Thus, the resulting causal ef-
fect amounts to about four percentage points. According to the result of
the two tailed t–test, this effect is statistically significant different from zero,
even according to the bootstrapping. As mentioned above, the result shows
that the ordinary t–statistic is biased downwards.
Eastern German firms which receive public R&D funds achieve on average
higher firm spec ific R&D intensities compared to firms that do not receive
public R&D support, given that the firms from both groups do not differ
with respect to exogenous variables that influence the probability of receiv-
ing public R&D subsidies. The results confirm that public R&D schemes
23
Table 4: Causal effect — firm sp ecific R&D–intensity
subsidized non–subsidized causal effect test statistic
firms
ˆ
E[Y
1
|S = 1] firms
ˆ
E[Y
0
|S = 0]
ˆ
θ
1
t–value
firms (per cent) (per cent) (percentage points) (bootstrap t–value)
622 6.57 2.63 3.94 8.24
∗
(5.32
∗
)
Note:
∗
indicates statistical significance in a two–tailed t–test at the 1% level.
in Eastern Germany are an important factor for stimulating private R&D
efforts.
The significantly higher R&D intensities for subsidized firms indicate that
complete substitution of public means does not take place, i.e. the absence
of perfect crowding–out. The recipients increase instead their private R&D
efforts in the case of public subsidization. This is especially important in a
transition economy like Eastern Germany, where private R&D is indispens-
able for creating innovative and viable economic structures after more than
40 years of a planned economy.
Of course, it be would interesting to know how large the net effect of public
funding is for the Eastern German manufacturing sector at all. The MIP
provides weights for its sampled firms which allow to calculate population
weighted descriptive statistics and, in our case, to estimate a macroeconomic
effect roughly. According to these information, the total R&D expenditure
in the Eastern German manufacturing sector in 1998 was about 3.84 billion
DM. Firms that participated in any public innovation scheme spent almost
3.4 billion DM of this amount. According to the result displayed in Table
4, we assume that 60% of recipients’ R&D activities are on average due to
public funding. Applying this rule of thumb, we derive a macroeconomic
effect of 2.04 billion DM according to subsidies. This effect is large com-
pared to other studies cited in section 2. However, keeping in mind that the
transformation process in Eastern Germany is heavily fostered by the gov-
ernment, this figure seems to be plausible. Of course, it would be desirable
to carry out a cost benefit analysis, but unfortunately the German Federal
Government does not provide any information on how the 2 billion DM of
24
public funding dedicated to the business sector (BMBF 2000) are allocated
to Eastern and Western German firms.
7 CONCLUSIONS
This papers provides new evidence to the discussion on whether public R&D
funds crowd out private investment in innovations. It is analyzed whether
the participation in public R&D programs leads on average to a higher R&D
intensity at the firm level. Using a non–parametric matching approach, we
compare the potential outcome of this group to a matched control group of
non–subsidized firms.
The analysis has some advantages over previous studies. The information
collected in the Mannheim Innovation Panel is not restricted to a particu-
lar measure but covers all public funding activities by the EU, the federal
government and the federal states in the years after reunification. However,
it is not possible to track in which program a firm participated with the
available information. The procedure used to identify the causal effect of
public R&D schemes is also new to this kind of literature. We use a non–
parametric matching approach to define a suitable control group.
The study comes up with following results: the causal effect identified is
significantly positively different from zero, i.e. firms that received public
funding achieve on average a higher R&D intensity than firms belonging
to the selected control group. The causal effect amounts to about four
percentage points on average. For example, a subsidized firm with a turnover
of 100,000 monetary units would on average have invested 4,000 monetary
units less if it did not participate in public R&D schemes.
ACKNOWLEDGEMENTS
We are grateful to the members of MIP team for providing the data and
to Fran¸cois Laisney, two anonymous referees and the associate editor of
the JBES for helpful comments. All remaining errors are, of course, the
responsibility of the authors alone.
25
References
Almus, M. (2001), Evaluating the Impact of Public Start-up Assistance –
Results from an Econometric Approach, ZEW Discussion Paper 01-23,
Mannheim.
Almus M., J. Egeln, M. Lechner, F. Pfeiffer and H. Spengler (1999),
Wirkungen gemeinn¨utziger Arbeitnehmer¨uberlassung in Rheinland–
Pfalz, Beitr¨age zur Arbeitsmarkt- und Berufsforschung 225, N¨urnberg.
Angrist, J., G. Imbens and D.B. Rubin (1996), “Identification of Causal Ef-
fects Using Instrumental Variables,” Journal of the American Statistical
Association 91, 444-455.
Ashenfelter, O. (1978), “Estimating the Effect of Training Programs on
Earnings,” Review of Economics and Statistics 60, 4-57.
Ashenfelter, O. und D. Card (1985), “Using the Longitudinal Structure of
Earnings to Estimate the Effect of Training Programs,” Review of Eco-
nomics and Statistics 67, 648-660.
BMBF (German federal ministry for education and research) (2000), Bun-
desbericht Forschung 2000, Bonn.
Busom, I. (2000), “An empirical Evaluation of the Effects of R&D Subsi-
dies,” Economics of Innovation and New Technology 9(2), 111–148.
Cochran, W. and D. Rubin (1973), “Controlling Bias in Observational Stud-
ies,” Sankyha 35, 417–446.
David, P.A., B.H. Hall und A.A. To ole (2000), “Is public R&D a complement
or substitute for private R&D? A review of the econometric evidence,”
Research Policy 29, 497–529.
Dehejia, R.H. and S. Wahba (1999), “Causal effects in nonexperimental
studies: reevaluating the evaluation of training programs,” Journal of
the American Statistical Association, 94, 1053-1062.
Ebling, G., C. Hipp, N. Janz, G. Licht and H. Niggemann (1999), “Inno-
vationsaktivit¨aten im Dienstleistungssektor — Ergebnisse der Innova-
tionserhebung 1997,” in: N. Janz and G. Licht (eds.), Innovationsak-
tivit¨aten in der deutschen Wirtschaft, Baden–Baden.
Efron, B. and R.S. Tibshirani (1993), An introduction to the bootstrap, Ne w
York.
26
Eurostat and OECD (1997), OSLO–Manual — Proposed guidelines for col-
lecting and interpreting technological innovation data, Paris.
Fitzenberger, B. und H. Prey (1998), “Besch¨aftigungs- und Verdienstwirkun-
gen von Weiterbildungsmaßnahmen im ostdeutschen Transformation-
sprozess: Eine Methodenkritik,” in: Pfeiffer, F. und W. Pohlmeier
(eds.), Qualifikation, Weiterbildung und Arbeitsmarkterfolg, Baden-
Baden, 39-96.
Greene, W.H. (2000), Econometric Analysis, New Jersey.
Harvey, A.C. (1976), “Estimating Regression Models with Multiplicative
Heteroscedasticity,” Econometrica 44, 461–465.
Hausman, J. (2001), “Microeconometrics,” Journal of Econometrics 100,
33-35.
Heckman, J.J. und V.J. Hotz (1989), “Choosing Among Nonexperimental
Methods for the Estimating the Impact of Social Programs: The Case
of Manpower Training,” Journal of the American Statistical Association
84, 862-880.
Heckman, J.J., H. Ichimura, J. Smith und P. Todd (1996), Characteriz-
ing Selection Bias using Experimental Data, mimeo, revised version is
published in Econometrica 66, 1017–1098.
Heckman, J.J., R.J. LaLonde und J.A. Smith (1999), “The Economics and
Econometrics of Active Labor Market Programs,” in: Ashenfelter, O.
und D. Card (eds.), Handbook of labor economics 3a, Amsterdam, 1865-
2097.
Hujer, R., K.-O. Maurer and M. Wellner (1997), The Impact of Training
on Unemployment Duration in West Germany, Discussion Paper 74,
Department of Economics, Johann Wolfgang Goethe-University, Frank-
furt.
Janz, N., G. Ebling, S. Gottschalk and H. Niggemann (2001), “The
Mannheim Innovation Panels (MIP and MIP–S) of the Centre for Eu-
ropean Economic Research,” Schmollers Jahrbuch – Journal of Applied
Social Science Studies 121, 123–129.
Klette, T.J., J. Møen und Z. Griliches (2000), “Do Subsidies to Commercial
R&D Reduce Market Failures? Microeconometric Evaluation Studies,”
Research Policy 29, 471–495.
27
Lach, S. (2000), “Do R&D Subsidies Stimulate or Displace Private R&D?
Evidence from Israel,” NBER Working Paper 7943, Cambridge, MA.
LaLonde, R.J. (1986), “Evaluating the econometric evaluations of training
programs with experimental data,” The American Economic Review,
76, 604-619.
Lechner, M. (1998), Training the East German Labour Force – Microecono-
metric Evaluation of Continiuous Training after Unification, Heidel-
berg.
Lechner, M. (1999), “Earnings and Employment Effects of Continuous Off-
the-Job Training in East Germany after Unification,” Journal of Busi-
ness and Economic Statistics, 17, 74-90.
Lichtenberg, F. (1984), “The Relationship between Federal Contract R&D
and Company R&D,” American Economic Review 74, 73–78.
Ragnitz, J. (2000), “Die Zukunft der Ost–F¨orderung,” Wirtschaftsdienst
2000/IV, 225–229.
Rosenbaum, P.R. and D.B. Rubin (1983), “The Central Role of the Prop en-
sity Score in Observational Studies for Causal Effects,” Biometrika 70,
41–55.
Roy, A.D. (1951), “Some Thoughts on the Distribution of Earnings,” Oxford
Economic Papers 3, 135–146.
Rubin, D.B. (1974), “Estimating Causal Effects of Tre atments in Random-
ized and Non–Randomized Studies,” Journal of Educational Psychology
66, 688–701.
Rubin, D.B. (1977), “Assignment to Treatment Group on the Basis of a
Covariate,” Journal of Educational Statistics 2, 1–26.
Silverman, B.W., 1986, Density Estimation for Statistics and Data Analysis,
London.
Stiglitz, J.E. and A. Weiss (1981), “Credit Rationing in Markets with Im-
perfect Information,” The American Economic Review 71, 393–410.
Verbeek, M. (2000), A Guide to Modern Econometrics, New York.
Wallsten, S.J. (2000), “The Effects of Government–Industry R&D Programs
on Private R&D: The Case of the Small Business Innovation Research
Program,” RAND Journal of Economics 31(1), 82–100.
28
APPENDIX: INDUSTRIES IN SAMPLE
Table 5: C lassification of Industry Dummies
Industry Dummy Description
industry 1 Food and beverages
industry 2 Textiles, clothes and leather goods
industry 3 Wood, paper, publishing and printing
industry 4 Fuels and chemicals
industry 5 Rubber and plastic products
industry 6 Non–metallic mineral products
industry 7 Basic and fabricated metals
industry 8 Machinery and equipment
industry 9 Office and communication equipment, electrical
machinery and components
industry 10 Medical and optical instruments
industry 11 Motor vehicles and other transport equipment
industry 12 Furniture products and n.e.c.
29
