Productivity Implications of R&D, Innovation, and Capital

Accumulation for Incumbents and Entrants: Perspectives from

a Catching-up Economy∗
Jaan Masso†, Amaresh K Tiwari ‡
arXiv:2205.10540v1 [econ.GN] 21 May 2022

Abstract

We study the productivity implications of R&D, capital accumulation, and innovation output
for entrants and incumbents in Estonia. First, in contrast to developed economies, a small
percentage of firm engage in formal R&D, but a much larger percentage innovate. Second,
while we find no difference in the R&D elasticity of productivity for the entrants and incum-
bents, the impact of innovation output — many of which are a result of ‘doing, using and
interacting’ (DUI) mode of innovation — is found to be higher for the entrants. Entrants who
innovate are 21% to 30% more productive than entrants who do not; the corresponding figures
for the incumbents are 10% to 13%. Third, despite the adverse sectoral composition typical
of catching-up economies, Estonian incumbents, who are the primary carriers of ‘scientific and
technologically-based innovative’ (STI) activities, are comparable to their counterparts in de-
veloped economies in translating STI activities into productivity gains. Fourth, while embodied
technological change through capital accumulation is found to be more effective in generating
productivity growth than R&D, the effectiveness is higher for firms engaging in R&D. Finally,
our results suggest that certain policy recommendations for spurring productivity growth in
technologically advanced economies may not be applicable for catching-up economies.

Keywords: Productivity, Entrants, Incumbents, R&D, Technological Innovations, STI & DUI

JEL Classifications: O31, O32

∗
This work was supported by the European Union Horizon 2020 Research and Innovation Action under
Grant Number 822781 and Estonian Research Council under Grant number PRG791. We owe thanks to
Statistics Estonia for their indispensable help in supplying the data. The authors also acknowledge support
for the compilation of the datasets used in the paper from the Estonian Research Infrastructures Roadmap
project ‘Infotechnological Mobility Observatory (IMO).’ We thank the participants of the 5th RIE Conference
and the 12th Conference on Model-based Evidence on Innovation and Development (MEIDE) Maastricht,
12/2021, and the participants of the Baltic Economic Association Conference, Tartu, 06/2021. Thanks are
due to Linas Tarasonis and Priit Vahter for providing valuable comments. The usual disclaimers apply.
†
Jaan Masso: University of Tartu, [email protected]
‡
Amaresh K Tiwari: University of Tartu, [email protected] & [email protected]
1 Introduction
The decline in productivity growth and job reallocation coupled with rising intra- and inter-
industry dispersion in productivity during the last two decades in the major advanced eco-
nomies have, as pointed out by Akcigit and Ates (2021), has coincided with decline in firm
entry rate and young firms’ share in economic activity. Since creative destruction — the pro-
cess by which newly innovated technologies replace older technologies — plays a central role
in many theories of growth, and since it is mainly attributed to innovation activities of fast
growing, young innovative firms, the decline in productivity growth and economic activities
of the young firms have prompted many recent studies on the innovative behaviour of the
entrants and incumbents.
While there are many such studies for the developed economies — a partial list includes
Czarnitzki and Delanote (2012), Coad et al. (2016), Acemoglu et al. (2018), Garcia-Macia
et al. (2019), Lubczyk and Peters (2020), and Akcigit and Ates (2021) — studies on innovat-
ive behaviour of entrants and incumbents in catching-up economies in Central and Eastern
Europe (CEE) are lacking. Studying the innovative behaviour of these groups of firms in
catching-up economies, which are at some distance from the technological frontier, is im-
portant not only because their National Innovation System (NIS) — the web of linkages that
direct flow of technology and information among actors such as firms, universities and govern-
ment research institutes — is distinct from those of the developed economies (Radosevic 1998;
Castellacci and Natera 2013; Cirillo et al. 2019),1 but also because policy prescriptions, as we
show, for spurring growth in developed economies may not be applicable to the catching-up
economies. Also, assessing the productivity implications of the firms’ innovative behaviour
can inform how the two groups of firms contribute to the aggregate productivity growth.
An outcome of Estonia’s — a representative CEE catching-up economy — distinctive
NIS vis-à-vis developed economies is that, in our sample, about 30% of the firm-years have
positive R&D expenses, but 74% of the firm-years have innovated to introduce at least one
new product in the market or a new process. This suggests that in Estonia more firms innovate
through the ‘doing, using and interacting’ (DUI) mode of innovation rather than ‘scientific
and technologically-based innovation’ (STI) (see Jensen et al. 2007). Moreover, as argued in
Peters et al. (2017), while investments in R&D increase the probability of realising product
or process innovations, R&D investment is neither necessary nor sufficient for firm innovation.
Focusing on R&D alone will, therefore, give a partial picture of the productivity implications
of innovative activities (see also Radosevic and Yoruk 2018, on the inadequacy of existing
measures, that are overly R&D oriented, for gauging technological change in middle-income
economies). We, therefore, in contrast to some of the papers mentioned above, develop an
empirical strategy to simultaneously estimate (1) the productivity elasticity of own R&D for
the entrants and the incumbents, and (2) the impact of technological — product and process
— innovations on productivity. Also, we pay particular attention to complementarity between
product and process innovation for the two group of producers.
1
Notwithstanding the various theorizations of the NIS and the continuing work to develop a more robust
concept of NIS (Cirillo et al. 2019), in various studies, such as Castellacci and Natera (2013) and Cirillo et al.
(2019), catching-up economies of the Central and East Europe are categorized separately from the Western
European countries and the US.

1
 Various explanations — not limited to those proffered by innovation studies — for the phe-
nomenon of declining productivity growth and job reallocation have been proposed. Czarnitzki
and Delanote (2012) and Coad et al. (2016) point that EU start-ups face higher entry and
growth barriers, including financial constraints, than their US counterparts. Akcigit and Ates
(2021) argue that decline in knowledge diffusion is likely to increase market concentration,
which implies that a new entrant is much more likely to compete against a dominant market.
This, they argue, will discourage new firm creation and is likely to result in decreased eco-
nomic activity among young firms. Acemoglu et al. (2018) claim that industrial policies that
subsidize (often large) incumbent firms, may also reduce economic growth by slowing down
reallocation and even discouraging innovation by both continuing firms and new entrants.
Acemoglu et al. (2018), who have a clear policy formulation for enhancing growth, build
on models of endogenous technological change, models of firm-level innovation (Klette and
Kortum 2004) and incorporate major elements from the reallocation literature2 to construct
a model of firm innovation and growth that enables an examination of the forces that jointly
drive innovation, productivity growth, and reallocation. In their model, incumbents and
entrants invest in R&D in order to improve over (one of) its many products. However,
firms are heterogeneous (high and low types) in their efficiency with which they innovate.
Firms that enter the market are disproportionately of high-quality type but may become low-
quality firms with a certain transition probability over time. This heterogeneity introduces a
selection effect, with concomitant reallocation occurring with the movement of R&D resources
(skilled workers) from less efficient innovators (struggling incumbents) towards more efficient
innovators (new firms). To promote growth and welfare they, therefore, propose taxes on the
continued operation of incumbents combined with a small incumbent R&D subsidy.
Aghion and Jaravel (2015) critiquing Acemoglu et al. (2018) point that their model, in
which R&D investments interact with general equilibrium effects, does not incorporate the
notion of absorptive capacity and that it would be fruitful for future research to do so. Aghion
and Jaravel (2015), referring to Cohen and Levinthal (1989), state that R&D not only cre-
ates new knowledge but also facilitates learning and building absorptive capacity. This is
particularly relevant for catching-up economies such as the Central and Eastern European
(CEE) countries, which lag behind the technological frontier, and where investing in absorpt-
ive capacity through R&D and better education can improve the ability to innovate and/or
imitate leading edge technologies (Aghion et al. 2011; Radosevic and Yoruk 2018). Griffith
et al. (2004) show empirically that (a) R&D affects both the rate of innovation and technology
transfers, and therefore failing to take into account R&D-based absorptive capacity results
in large underestimates of the social rate of return to R&D, and (b) countries and industries
lagging behind the productivity frontier catch-up particularly fast if they invest heavily in
2
This literature emphasizes selection mechanisms, which characterise industries as collections of firms that are
heterogeneous in their productivity, and link firm productivity levels to their performance and survival in the
industry (Jovanovic 1982; Hopenhayn 1992; Asplund and Nocke 2006). This heterogeneity induces a selection
effect, by which reallocation of market shares to more efficient producers, either through market share shifts
among incumbents or through entry and exit, drive aggregate productivity movements. Low productivity
plants are less likely to survive and thrive than their more efficient counterparts, creating selection-driven
aggregate (industry) productivity increases. Decker et al. (2014) document that, given initial size, more
productive firms grow faster than the less productive ones; however, the contributions entrants make to
growth are considerably heterogeneous.

2
R&D. These also imply that many of the policy recommendations in Acemoglu et al. (2018)
for the US and the technologically advanced economies may not be applicable for catching-up
economies.
Parisi et al. (2006) show that fixed capital spending by Italian medium-low and low-tech
firms increases the likelihood of introducing a process innovation. This, as Parisi et al. (2006)
argue, suggests that physical capital stock embodies technological progress. In catching-up
economies (e.g. Estonia) — (a) which tend to grow more due to the imitation activities of
the firms, and (b) where a higher share of firms are in medium-tech and low-tech sectors
— one could, therefore, expect capital stock accumulation to have significant productivity
implications.
Lubczyk and Peters (2020)(LP hereafter) state that little scholarly research has assessed
the dynamics studied in Acemoglu et al. (2018) with firm level empirical applications. Czarnitzki
and Delanote (2012) using Flemish data show that young innovative companies grow faster
than the other, already fast-growing firms. Coad et al. (2016) using Spanish data show that
young firms — in terms of sales, productivity and employment growth — benefit more from
R&D, but R&D investment by young firms is riskier than R&D investment by more mature
firms. LP using a novel German data set find that entrants experience significantly larger
gains from investments in R&D than incumbents and that returns to R&D for entrants are
considerably more heterogeneous than that for incumbents. Besides, LP’s is a comprehensive
study of regional spillover effects for the two groups of firms.
To assess the relevance of the above discussed theories for a catching-up economy as
Estonia, the main objectives we formulate are:

(I) Given that reallocation, particularly due to entry and exit, accounts for a substan-
tial proportion of productivity growth and since this reallocation, both of production
and R&D inputs, is to a significant extent due to productivity impacts of innovation
undertaken by heterogeneous incumbents and entrants, we study the productivity im-
plications of (a) R&D, (b) knowledge spillover, and (c) technological innovations for
incumbents and entrants.

(II) Since catching-up economies have been known to grow by engaging in imitation/learning
activities (Aghion et al. 2011; Radosevic and Yoruk 2018), embodied technological
change through capital accumulation is likely to have a significant productivity im-
pact. We, therefore, compare the productivity impact of capital accumulation with the
productivity impact of R&D for both groups of producers.

(III) Our third objective is to compare the productivity growth impacts of R&D and capital
stock for Estonia, a catching-up country, to those in LP’s, who have undertaken a similar
analysis for Germany, a more technologically advanced country, and understand the
sources of the difference in the impacts. LP’s, to our knowledge, is the most recent work
that has comprehensively looked at productivity impact of R&D and regional spillovers
for the two groups of producers in a developed economy. While the comparison puts
our results in perspective, we believe it will also inform better innovation and growth
enhancing policies for Estonia and catching-up economies.

3
 Before proceeding further, we document some of our findings. To begin with, though all
firms gain from investing in R&D, we do not find any difference in the average productivity
elasticity with respect to R&D for the incumbents and the entrants. Our results therefore
do not match the findings in Coad et al. (2016), Acemoglu et al. (2018) and LP, where the
impact of R&D on productivity is much higher for entrants as compared to the incumbents’.
However, we do find that impact of technological innovation — many of which are a result
of DUI mode of innovation — on productivity is much higher for the entrants as compared
to the incumbents. We explored the sources of the productivity gain for both the groups
of producers, and found that the gain is largely due to complementarity between product
and process innovation; innovating entrants, however, realize a much higher complementarity
than incumbents.
Now, while Estonian entrants were found to be less able than their counterparts in Ger-
many in translating R&D into productivity gains, we found an average Estonian incumbent
to be on par with an average German incumbent. This is despite the adverse sectoral compos-
ition of the Estonian economy, where, as is typical of catching-up CEE economies, compared
to an average European Union member state, there are fewer firms and fewer people employed
in sectors where the returns to R&D are highest. Moreover, because the bulk of STI based
innovative activities in Estonia is carried out by the incumbents, we find Estonian incumbents
to be playing an important role as carriers of STI based innovative activities.
We find a robust evidence that embodied technological change through capital accumu-
lation is more effective in generating productivity growth than R&D expenditure. However,
since R&D or STI based innovative activities itself help build absorptive capacity, we also
find that productivity elasticity with respect to capital is higher for firms that report enga-
ging in formal R&D. Finally, in results related to spillover effects, we find it is mostly the
incumbents who benefit from intra- and inter-industry knowledge that is generated by other
firms — primarily by the incumbents — situated in close proximity. However, the spillover
effects are higher for firms that do not reportedly engage in formal R&D, suggesting that
these firms, too, work on building absorptive capacity.
The remainder of this paper is structured as follows. Section 2 reviews related literature
and in section 3 we describe the data used for our study. In section 4, we explain the empirical
strategy employed in the paper. In Section 5, we present and discuss the empirical results,
while section 6 draws concluding remarks. Certain details of the econometric methodology
have be relegated to the appendix.

2 Literature Review
In subsection 2.1 we discuss some literature on the relationship between firm age, innovation
and productivity, while in subsection 2.2 we review some relevant studies on innovation,
reallocation, and productivity growth for the CEE countries.

2.1 Innovation among Entrants and Incumbents

Among the large strand of literature on how different firms contribute to productivity and
productivity growth, many studies have examined how dynamics between entrants and in-

4
cumbents impact aggregate productivity development, highlighting both differential and in-
terrelated effects. In this section, we review some of the literature on the innovative behaviour
of the two groups of firms.
Entrants and incumbents have been described as two distinct groups of firms (Berchicci
and Tucci 2009; Lubczyk and Peters 2020). Entrants wanting in experience, still need to
learn about the economic environment in which they operate. Incumbents, on the other
hand, have accumulated considerable experience in their competitive environment and com-
mand well established capabilities. Entrants, however, are seen as being more expeditious
than incumbents due to lack of structural inertia to reorganisation, faster decision-making
processes, streamlined operations, and targeted innovation. These result in a timely response
to changing industry environments, and also make them more efficient innovators compared
to incumbents.
However, entrants are often financially constrained, and investing a large part of the
limited resource endowment in R&D activities, which are inherently uncertain, can pose a
significant business risk. Lack of experience and limited expertise may create further com-
plications for the innovation process, especially when unforeseen circumstances arise. On
the other hand, however, entry is envisaged as the way in which firms explore the value of
new ideas in an uncertain context, and that entry, the likelihood of survival, and subsequent
conditional growth are determined by barriers to survival (Audretsch 1995; Huergo and Jau-
mandreu 2004). In this framework, entry is innovative and occurs at a higher rate at the
start of an industry when uncertainty is high (Klepper 1996); the likelihood of survival is
lower the higher the risk; and the growth from successful innovation is higher the higher the
barriers to survival. Successful innovation, therefore, is likely to result in substantial relative
productivity growth for newly established enterprises. Given that most young firms are small,
successful innovation can often disproportionately spur growth and contribute to substantial
increases in employment, revenue and future profitability (Haltiwanger et al. 2013). There-
fore, one would expect the impact of R&D spending on productivity to be volatile for the
new entrants.
Incumbents, having survived through time and participated longer in the market, and
have many factors working in their favour as far as innovative activity is concerned. First,
even though incumbents often lack the organisational agility of smaller and younger com-
petitors, they may compensate for this with resources — such as financial and marketing
capabilities — and innovation capacity built over time (Berchicci and Tucci 2009). Moreover,
an incumbent can build on its existing infrastructure even as existing business experience
and infrastructure may enable the incumbent to pursue more ambitious R&D projects. Also,
the experience of having conducted successful innovation in the past increases the likelihood
of future innovation (Peters 2009; Raymond et al. 2010) and may help such organisations to
achieve higher levels of efficiency in carrying out their R&D activities (Lööf and Johansson
2014).
The nature of innovations, too, is often different for the two groups of producers. Since
incumbents would like to safeguard their profits from the established products and production
technologies in place, their innovative activity is therefore more often of an incremental
nature, whereas young firms, in order to create higher quality products and overtake product

5
lines previously operated by incumbents, are more inclined to exploit new ideas and engage
in radical innovation (Acemoglu and Cao 2015). Besides, radical innovation often entails
costly organisational restructuring, which may deter incumbents from undertaking radical
innovation (Berchicci and Tucci 2009).

2.2 Relevant Studies on Innovation and Reallocation in CEE Countries

While there are many studies on innovation, reallocation, and productivity growth for the US
and Western European countries, there are only a few as far as Central and East European
(CEE) countries are concerned. Studies on productivity growth due to selection and realloc-
ation for the CEE countries are by Masso et al. (2004) and Bartelsman et al. (2013). Masso
et al. (2004) find that newly formed firms had a higher survival rate than incumbents, and
that the reallocation of production factors, especially due to the exit of low productivity
units, contributed to the productivity growth. Bartelsman et al. (2013) find that, although
the covariance between firm-size and productivity, a measure of resource misallocation, is low
in Eastern Europe, it has been increasing substantially over the last couple of decades.
Masso and Vahter (2008) point out that to sustain initial growth rates during the trans-
ition period in CEE countries, which was based on initial capital accumulation and imitation
of technologies applied elsewhere, these countries will need to rely increasingly on their own
innovation. Due to their attempts to establish knowledge-based economies and to increase
business R&D, there is growing interest in studying the relationship between innovation, pro-
ductivity and growth in the CEE countries. Among the few studies that have studied the
productivity impact of R&D in a CEE country is one by Liik et al. (2014), who estimate
elasticities using industry level data to conclude that in comparison to OECD economies,
R&D investments play a relatively limited role in determining the productivity and efficiency
levels of Estonian industries. Lacasa et al. (2017), studying the technological capabilities of
CEE economies based on patent data, find that the CEE economies reduced their technolo-
gical activities drastically after 1990, and that the recovery of CEE economies with respect
to technological capabilities is unfolding very slowly. They find that CEE countries innovate
in less dynamic technological sectors and contribute only to a limited number of fields with
growing technological opportunities.
Recent studies, such as Bruno et al. (2019), find that while R&D intensity has been
effective in closing the distance to the productivity frontier, R&D embedded in purchased
equipment and machinery have played an important role in reducing the distance. Filippetti
and Peyrache (2015), studying the role of the technology gap in explaining labour productivity
differences in 211 European regions over the years 1995—2007, find that labour productivity
growth in CEE (or lagging behind) regions, productivity growth is mainly driven by capital
accumulation.

3 Data and Variables: Definitions and Description

For our study, we use eight waves of the Estonian Community Innovation Survey (CIS),
CIS2004 to CIS2018, which is a biennial survey about innovation activities in Estonian en-
terprises. For a detailed description of the CIS data and the usage of the data for scholarly

6
analysis of various issue related to innovation, see Mairesse and Mohnen (2010). Firm bal-
ance sheet information containing profit and loss statements was obtained from the Estonian
Business Registry, which is a census data. Wherever possible, the missing information in the
Business Registry data was obtained from EKOMAR, a survey data.
One of the key variables from the CIS surveys used for our analysis is the R&D expenditure
of firms. However, as non-innovative firms — firms that have not innovated any product or
process, or have no unfinished innovative activities — in the CIS data are not required to
report their R&D expenses, we do not observe R&D expenditure for the non-innovative firms.
The CIS questionnaire, however, allows us to distinguish between ‘potentially’ innovative
firms and firms that do not intend to engage in R&D among the non-innovative firms (see
Savignac 2008). This distinction is made on the basis of firm responses to the question
on factors that might have thwarted their R&D and innovative activities. We classify non-
innovative firms who face such factors as potentially innovative, whereas non-innovating firms
that do not face any of these factors are classified as firms that do not intend to engage in
R&D activities. For these non-innovators that do not wish to engage in R&D activities, it
can be safely assumed that they have no R&D expenses.
Since we do not observe R&D expenditure of potentially innovative firms, we drop such
firms from our analysis. However, non-innovative firms that do not wish to engage in R&D
activities are retained. This, as discussed in the next section, allows for us to estimate the
spillover effects of external knowledge stock as well as the impact of product and process
innovation on productivity for all firms: firms that have and that do not have R&D expenses.
After further removing firms with missing data, our sample comprised of 7,586 firm-year
observations from 3,068 firms. The minimum number of observations per firm is 1, the
maximum, 8, and the average number is about 2.5 years.
Since one of our aims is to compare our findings to those in LP, we, as in LP, define
entrants as firms that are new to the market and have been active for eight or less years,
while incumbents are those that have been active for more than eight years. Our set of
entrants excludes firms resulting from mergers, break-ups, split-off or restructuring of a set of
enterprises. Those resulting from change of activity, too, are excluded. There is no theoretical
basis for the above definition of an entrant; in the literature, an entrant’s maximum age has
ranged from three to ten years. Of the 7,586 firm-year observations, 6,086 are incumbents and
1,500 are entrants. Estonian firms are relatively young, where the average age of incumbents
is 17 years, while that of entrants is 5.5 years.
Also, we define a firm, j, in year, t, to be an innovator, Ijt = 1, if the firm had introduced
new products, PjtD = 1 or processes, P C = 1, to the market. Value-added productivity is
jt
our dependent variable and performance outcome measure. It is the ratio of value-added,
the difference between sales revenues and the value of intermediate inputs, to the number of
employees.
In order to convert the book value of the gross capital stock into its replacement value, we
use the perpetual inventory method described in Salinger and Summers (1983). According
to this method, the replacement value of the capital stock is equal to the book value of fixed
assets for the first year the firm appears in the data. For the subsequent years, first, the

7
useful life of capital goods, Ljt , at time t is calculated as:

GKj,t−1 + Ijt
Ljt = ,
DEP Rjt

where GKj,t−1 is the reported value of gross property, plant, and equipment at time t − 1,
Ijt is the investment in the same for the period t, and DEP Rjt is the reported depreciation.
Then Ljt is averaged over time to obtain L, which is then used in the following formula to
obtain the replacement value of the capital stock of a firm in industry k:

pk
 
Kjt = Kj,t−1 kt + Ijt (1 − 2/Ljt ),
pt−1

where pkt is the deflator for industry, k. The second term represents the amount of capital
stock that depreciates each year and is based on the assumption that economic depreciation
is a double declining balance. For new firms and for existing firms that appear again after a
gap in later time periods in the data, the book value of the capital stock in the first year is
taken as the replacement value. However, for firms that after a gap reappear in later years,
this method will not yield as good an estimate of the replacement value as for the firms for
whom a long, continuous time-series is available.
In Table 1, we can see that the incumbents on average have significantly higher labour
productivity, larger capital stocks, higher investment and more employees. As far as innov-
ative behaviour of the firms is concerned, as can be seen in Table 1, a majority of the firms
(about 74%) in the estimation sample are innovative even as the percentage of firms that have
positive R&D is only about 30%. Secondly, incumbents have a higher propensity to invest in
R&D and are also likely to invest more in R&D than the entrants. However, it is the entrants
who are more innovative as far as technological — product and process — innovations are
concerned. Also, for the incumbents, log of R&D intensity increased substantially from an
average of 1.76 in the pre Financial Crisis years to an average 6.93 in the post Crisis years;
the corresponding figures for the entrants are, 2.03 and 6.95.
In Table 2, we can see that high-tech and medium-high-tech manufacturing, which are
most R&D intensive and most innovative, constitute about 13% of the firms in the sample.
Majority of the firms (70%) are in the low-tech and medium-low-tech manufacturing and
less knowledge intensive services; these sectors, though less R&D intensive, are still quite
innovative.
In our analysis we include a dummy for north Estonia, which takes value 1 if the company
is located in Harju county. Harju county is the biggest of all Estonian counties in terms
of population and economic activity, and includes the national capital city, Tallinn. The
rational for including this dummy is that firms located in Harju county, which qualifies as the
economic hub of Estonia, are likely to be better networked with implications for productivity
due to agglomeration. We find that a significantly higher proportion of entrants are based in
northern Estonia (see Table 1).
The external knowledge capital, Ejt , is intended to capture knowledge spillovers among
firms. In our empirical analysis, we differentiate between different types of spillovers: intra-

8
industry and inter-industry R&D spillovers to measure to what extent entrants and incum-
bents differ in their capacity to benefit from R&D knowledge that is available to the firm
within and outside its own industry.
The measure of the intra-industry knowledge capital of firm i in industry k (based on
two digit NACE Rev. 2 codes) in period t is the weighted sum of R&D expenditures per
employee of other firms in industry k:

X k
Rjt
Intra-Industry R&D = wij , (3.1)
j6=i
Lkjt

where Rjt is the R&D expenditure of firm j, Ljt is the number of employees employed by
the firm, and the weight, wij , is the inverse of the geographical distance between the capital
of the county in which firm i is located and capital of the county in which firm j belonging
to the same industry, k, is located; if firm j happens to be in the same county then wij is
taken as 1. The measure of the inter-industry knowledge of firm i in industry k (NACE two
digit) in period t is the weighted sum of R&D expenditures per employee of firms in other
industries:

XX l
Rjt
Inter-Industry R&D = wij , (3.2)
l6=k j
Lljt

where, again, the weight wij is the inverse of the geographical distance between the capital of
the county in which firm i is located and the capital of the county in which firm j belonging
to a different industry, l, is located.
In Table 1 we can see that incumbents on average are based in or based close to re-
gions/counties with higher average external knowledge; the differences hold for both within
and across industry comparisons. Here, we would like to note that even at this broader
definition of industry (NACE 2-digit), about 2% of the firm-year observations had zero intra-
industry knowledge flows. This is because a relatively smaller number of firms, about 30%,
in the estimation sample invested in R&D.

[Table 1 about here]

[Table 2 about here]

4 Empirical Strategy
For firm j = 1, . . . , J and time t = 1, . . . , T , we observe revenue (Yjt ), labour (Ljt ), capital
(Kjt ), material inputs (Mjt ), R&D expenses (Rjt ), which can also be zero, and (Ejt ), which
is some measure of external knowledge capital.
Now, while a measure of R&D capital stock would be preferred as an innovation input,
we are unable to estimate the knowledge capital from the past R&D expenditures using the
perpetual inventory method. This is because CIS surveys are conducted once every two
years and, secondly, we do not have a balanced panel data. Instead, following Crépon et al.
(1998) and more recently Raymond et al. (2015) and Baumann and Kritikos (2016), we proxy

9
knowledge capital using current R&D expenditure, Rjt . This proxy is based on the implicit
assumption that firm R&D investments are strongly correlated (and roughly proportional)
to their R&D capital stock measure, and that R&D engagement and intensity persists over
time.3
Let ωjt be the firm j and time t specific total factor productivity (TFP), which is ob-
served by the firms but not by the econometricians. For notational convenience, we drop
the subscript, jt, as of now. For estimation, we employ the “control function” methodology
developed by Ackerberg et al. (2015) (ACF), who consider a value-added production function.
The value-added production function, as ACF explain, can be derive from the gross output
production function that is Leontief in material inputs, M . So, when R&D expenditure is
positive, i.e. R > 0, output is given by

Y = min{BLβl K βk Rβr E βe eω , βm M }.

Y −M
The above implies that the value-added productivity, , when R > 0 can be written as
L
 β k  β r
Y −M 1 βl +βk +βr −1 K R
= (1 − )BL E βe eω . (4.1)
L βm L L

As discussed in the section on data, not all firms, innovating or otherwise, have positive
R&D expenditure. We nonetheless keep such firms in our analysis for two reasons. First,
this allows us to estimate the spillover effects of external knowledge capital for firms that
do not engage in R&D. Second, as discussed below, we are able to use the information on
product and/or process innovation, even for firms that do not have positive R&D expenses,
to endogenise the evolution of ω and assess the implications of innovation on productivity,
ω. The Leontief production function when R&D expenditure, R, is equal to zero and when
material input is proportional to output is given by

0 0 0
Y = min{B0 Lβl K βk E βe eω , βm
0
M },

which implies that the value-added productivity when R = 0 is given by

 β 0
Y −M 1 βl0 +βk0 −1 K
k 0
= (1 − 0 )B0 L E βe eω . (4.2)
L βm L

Taking logarithm of the value-added production functions when R > 0 and when R = 0 and
combining the two we get

y =(1 − D 0 )(β + βl+ l + βk k + βr r + βe e) + D 0 (β 0 + βl0+ l + βk0 k + βe0 e) + ω + ǫ, (4.3)

where y is the natural logarithm of the value-added productivity and D 0 is a dummy variable
that takes the value 1 when R&D expenditure is zero. The term βl+ = βl + βk + βr − 1 and
βl+0 = βl0 + βk0 − 1. The lower-case symbols, l, k, r, and e, on the RHS represent natural logs
3
With a big overlap of firms between the different waves in the Estonian CIS, we find some evidence of this
persistence.

10
 K R
of L, , , and E respectively. The constants β and β 0 are ln(B) and ln(B0 ) respectively.
L L
The term ǫ is the measurement error in value-added productivity.4 Since the specification in
(4.3) log-linear, the coefficients of the input variables can be interpreted as elasticities. As
is evident from the equation, we allow the productivity elasticities with respect to the input
variables to be different for firms that engage in R&D and firms that do not.
Since the unobserved productivity shocks, ω, are known to the firm when it makes its
input choices, we are confronted with the problem of endogeneity. To resolve the issue
of endogeneity, as stated earlier, we employ the ‘control function’ method by ACF. The
method requires that the outcome variable be value added, and depending on the timing of
investment and input decisions, it involves the use of economic theory to derive a ‘proxy’
for the anticipated productivity shock, ω. The proxy is obtained by assuming that ω can
be inverted out from certain firm inputs if the firm has adjusted these inputs optimally in
response to the ω it observed.
Now, in ACF the evolution of productivity, ωjt , has been assumed to be exogenous; that
is, productivity shocks evolve according to the first order Markov Process: p(ωj,t+1 |Ijt ) =
p(ωj,t+1 |ωjt ), where Ijt firm’s information set at t, which includes current and past pro-
ductivity shocks {ωjτ }tτ =0 but does not include future productivity shocks {ωjτ }∞
τ =t+1 . The
distribution, p(ωj,t+1 |ωjt ), which is stochastically increasing in ωjt , is known to the firm.5
Griliches (1979), however, points out that while R&D affects output, R&D is determined
both by past output and the expectations of future output. And thus past R&D efforts
and innovation output affect the evolution of productivity, ωjt , and the expectation of ωjt+1
affects the endogenous choice of Rjt . Doraszelski and Jaumandreu (2013)(DJ) endogenize
productivity evolution by including a measure of R&D investment: p(ωj,t+1 |ωjt , Rj,t ). Ac-
cording to this formulation, a firm by making investments in R&D alters the probability of
receiving a innovation, which in turn alters the distribution of productivity that it faces in
future periods.
Peters et al. (2017), endogenize productivity evolution by making ωjt depend on innova-
tion output instead of past innovation input. More specifically, productivity, ωjt, is modelled
D , P C ), where
to evolve according to the following controlled Markov process: p(ωjt |ωj,t−1 , Pj,t j,t
D is an indicator variable that takes the value 1 if firm j innovated at least one product
Pj,t
C is an indicator variable that takes the value 1 if firm j innovated
and 0 otherwise, and Pj,t
new processes in period t and zero otherwise. Here the assumption about the innovation
D and P C are a result of R&D efforts in the past. By including the
process is that both Pj,t j,t
innovation process in the model, rather than linking R&D directly to productivity as in DJ,
Peters et al. (2017) aim to gain additional insight into whether R&D improves productivity
through the demand side or the cost side of the firm’s operations. This is because product
innovations increase revenue by expanding the firm’s demand, whereas process innovations
increase revenue by improving the efficiency of scarce factors and thereby reducing firm’s cost
of production (see also Hall 2011).
4
See Ackerberg et al. (2015) for a discussion on the interpretation of ǫjt .
5
If only to mention, the transitory shocks, ǫjt , in (4.3) satisfy E[ǫjt |Ijt ] = 0.

11
 The Markovian assumption implies that

D C D C
ωjt = E(ωjt |ωj,t−1 , Pj,t , Pj,t ) + ξjt = g(ωj,t−1 , Pj,t , Pj,t ) + ξjt . (4.4)

In the above, productivity ωjt in period t has been decomposed into expected productivity,
D , P C ), and a random shock, ξ . While the conditional expectation function
g(ωj,t−1 , Pj,t j,t jt
D , P C ), ξ does not. The
g(.) depends on the already attained productivity ωj,t−1 and (Pj,t j,t jt
D , P C ). As discussed in
residual, ξjt , by construction is mean independent of ωj,t−1 and (Pj,t j,t
DJ, the productivity innovation, ξjt , represents the uncertainties that are naturally linked
to productivity and the uncertainties inherent in the innovation process such as degree of
applicability and success in implementation.
The formulation of productivity evolution in (4.4) (a) captures the lag dependence of
productivity, ωjt, and (b) by allowing innovations to affect the expected productivity, g(.),
and hence labour productivity, it captures the fact that R&D expenditures alone are not
sufficient to generate productivity improvements. This formulation, therefore, allows us to
estimate the productivity impact of technological innovations. The conditional expectation
function g(.) for the specification in (4.3) is estimated non-parametrically along with the
parameters of the production using the two step control function procedure proposed in
ACF.6
Since we are interested in estimating the impact of technological innovations on productiv-
ity for incumbents and entrants, we estimate the difference,

∆ = E[g1∨1 (ωj,t−1 , .)] − E[g0∧0 (ωj,t−1 , .)] (4.5)

for the entrants and the incumbents, where

D C D C
g1∨1 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 1 and/or Pjt = 1 and
D C D C
g0∧0 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 0 and Pjt = 0, (4.6)

and the expectations are taken over ωj,t−1 . The difference, ∆, in (4.5) is the average treatment
effect (ATE) of innovation on productivity; details on the identification of ATE can be found
in Appendix A.
To understand how (i) product innovation only, (ii) process innovation only, and (iii)
product and process innovations contribute to the ATE, ∆, of technological innovation we
estimate the following:

∆10 = E[g1∧0 (.)] − E[g0∧0 (.)],

∆01 = E[g0∧1 (.)] − E[g0∧0 (.)], and
∆11 = E[g1∧1 (.)] − E[g0∧0 (.)], (4.7)
6
We use the STATA’s ‘prodest’ command, which has been developed by Rovigatti and Mollisi (2018), to
implement the control function method in ACF. The endogenous() option allows users to specify one or more
variables that endogenously affect the dynamics of productivity, ωjt .

12
where

D C D C
g1∧0 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 1 and Pjt = 0,
D C D C
g0∧1 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 0 and Pjt = 1, and
D C D C
g1∧1 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 1 and Pjt = 1. (4.8)

Given the definitions of g1∧0 (.), g1∧0 (.), and g1∧0 (.) in (4.8) and g0∧0 (.) in (4.6), it is clear
that ∆10 in (4.7), for the matched sample of product innovators only and non-innovators, is
the ATE of product innovation only; ∆01 , the ATE of process innovation only; and ∆11 , the
ATE of product and process innovation.7 As shown in Appendix A, if

∆11 ≥ ∆10 + ∆01 , (4.9)

that is, if the impact of product and process innovation exceeds the sum of the impacts of
only product innovation and only process process innovation, then the product and process
innovation can be said to be complementary. We refer to Milgrom and Roberts (1990) for a
general treatment of complementarities between organizational decisions, and to Athey and
Schmutzler (1995) for complementarities between product and process innovation.

5 Results
First, in subsection 5.1, we discuss the results obtained from the estimation of the production
functions in equations (4.3), focusing on the average impacts of internal R&D investments
and capital accumulation on productivity for the entrants and incumbents. In this subsection,
we also compare our estimates of the impact of internal R&D and capital accumulation
on productivity with those estimated by Lubczyk and Peters (2020)(LP) for Germany. In
subsection 5.2 we study the differential spillover effects from knowledge that is produced
outside the firms’ boundaries. Finally, in subsection 5.3, we discuss the estimates of the
impact of technological innovation on productivity for the two group of producers.

5.1 Average Returns to Own R&D and Capital Accumulation for Entrants
and Incumbents
Using the empirical framework outlined in section 4, we first study how the productivity
effects of (1) innovation effort as proxied by R&D expenditure and (2) capital accumulation
differ between entrants and incumbent firms. In Table 3, we display the results of estimating
equation (4.3).
In Table 4 we compare the estimated elasticities of productivity with respect to own R&D
and capital stock with those estimated in LP for Germany. An important difference when
7
Ideally, ∆ in (4.5) is the weighted sum, ∆ = w10 ∆10 + w01 ∆01 + w11 ∆11 . The weights, w10 , w01 , and w11 ,
N10 N01 N11
respectively are , , and , where N10 is the number of firms that innovated only products, N01 is
N N N
the number of firms that innovated only processes, N11 is the number of firms that innovated both products
and processes, and N = N10 + N01 + N11 . However, if, for example, we use different matching criteria for
computing ∆11 and ∆ so that the sample of firms used for computing ∆11 is not contained in the sample of
firms used to compute ∆, then the equality may not hold.

13
compared to Germany is that in Estonia, a large percentage of innovating firms do not engage
in formal R&D. Estimating only the productivity elasticity with respect to R&D would,
therefore, give a partial picture of productivity implication of innovation. We, therefore, in
addition to estimating productivity elasticity with respect to R&D, as discussed earlier, in
subsection 5.3 estimate the impact of technological innovations on productivity along. This
required us to keep all firms in the estimation sample, and since the productivity elasticities
with respect to other inputs such as labour and capital could well differ for firms with positive
R&D (firms that engage in STI based innovative activities) and zero R&D, we estimated
separate productivity elasticities for firms with positive R&D and firms with zero R&D.
We find that the average productivity elasticity of R&D expenditures is significantly
positive for both entrants and incumbents. However, entrants’ elasticity of productivity with
respect to R&D does not differ from that of incumbents’ systemically across specifications
and methods. This is contrary to the findings in Acemoglu et al. (2018) and LP, where
entrants are disproportionally abler than incumbents in translating their R&D endeavours
into productivity gains.
Now, in our sample there are fewer firms in sectors that are R&D and knowledge intensive,
and where returns to R&D is high. In our sample, about 30% of the firm-years have positive
R&D, which is mostly concentrated in high-tech and medium-high-tech manufacturing (see
Table 2). However, in terms of value-added, we find that the average (averaged over the years)
share of high-tech manufacturing within manufacturing is less than 7%, and average share of
medium-high-tech manufacturing is about 22%. Though the shares have generally increased
over the year, they are lower than the corresponding EU27 2006 shares of 12.5% and 32%
respectively (see European Commission 2011, chapter 5, p. 121). Therefore, it could be due
what is known as the compositional effect (see Castellani et al. 2019) that our result is not
consistent with the theoretical argument that young firms that do not have a well established
product portfolio on the market benefit more from investing in R&D.
To examine whether, compared to the technologically advanced countries, the lower aver-
age elasticity with respect to R&D for Estonian entrants is due to the compositional effect, in
Table 5 we estimated separate elasticities for (a) firms in the high-tech and medium-high-tech
manufacturing and (b) firms in the remaining sectors. We find that entrants in high-tech and
medium-high-tech manufacturing are not particularly abler than entrants in other sectors in
translating R&D into productivity gains. Having ruled out compositional effect as a likely
cause for the lower average elasticity with respect to R&D, we take it as given that com-
pared to their counterparts in technologically advanced economies, Estonian entrants’ are
inefficient in STI mode of innovative activities. Further exploration of the causes behind the
comparatively lower elasticity with respect to R&D is outside the scope of the paper.
Incumbents in high-tech and medium-high-tech manufacturing, as we can see in Table 5,
do earn a significantly higher reward for investing a unit more of R&D than incumbents in
other sectors. In other words, had the sectoral composition of the Estonian economy been
similar to the technologically advanced countries or to the EU average, the average elasticity
with respect to R&D for Estonian incumbents would have been higher.
So, (a) despite the limitation posed by the sectoral composition, as can be seen in Tables 3
and 4, Estonian incumbents are no less abler than Estonia entrants and on par with German

14
incumbents in translating R&D efforts into productivity gains. Now, unlike in LP’s paper
for Germany, where more than half of the firms are entrants, and all of whom engage in
formal R&D, in our sample, about 20% of the firms are entrants,8 of which only 30% are
engaged in formal R&D. In other words, (b) the bulk of STI based innovation in Estonia is
carried out by Estonian incumbents. Given (a) and (b), we can, therefore, conclude that in
Estonia, incumbents play an important role as carriers of STI mode of innovation. STI mode
of innovation, which involves highly skilled scientific personnel, exploits possibilities at the
frontier of knowledge and, through innovations, pushes the frontier, too. Besides, also through
it’s interconnections with the DUI mode, the STI mode enhances the ability of economies to
import and adapt new technologies and innovations developed elsewhere (Jensen et al. 2007).

[Table 3 about here]

[Table 4 about here]

[Table 5 about here]

The coefficient of D0 is positive and significant for the baseline specification. This result,
however, is reversed once we allow for spillover effects by augmenting the baseline specification
with measures of external knowledge. Since, as discussed in the next subsection, the spillover
effects are particularly significant for the firms that do not engage in R&D, it is likely that D 0 ,
which is a dummy for zero R&D investment, is capturing the spillover effects in the baseline
specification. This then suggests that firms that do not engage in R&D are less productive.
We also find that both incumbents and entrants experience larger productivity gains from
additional capital accumulation than R&D expenditure, suggesting that in Estonia ‘embodied
technological change’ through capital accumulation, most of which is imported, has been
more effective in generating productivity growth (Castellani et al. 2019). Importantly, the
average elasticity with respect to capital for R&D performing firms was found to be higher
than for those firms that did not engage in R&D. For incumbents, the estimated elasticity
with respect to capital for R&D performing firms was found 1 to 3 percentage points higher
than for those firms that did not engage in R&D. As the test of equality of elasticity of
productivity with respect to capital for firms with and without R&D has been rejected (see
Table 3), these differences, though small, are significant. The fact that R&D performing
firms gain more from capital accumulation than non-R&D firms suggests that R&D helps
firms build absorptive capacity, which enables more efficient transfer of technology embodied
in capital good (Griffith et al. 2004; Aghion and Jaravel 2015).
Comparing elasticity of productivity with respect to capital for the two countries in Table
4, we find that, the elasticity is higher for the Estonian incumbents as compared to the German
ones. Now, Ortega-Argilés et al. (2014, 2015) find that in traditional low-tech industries,
which focus on process innovation, productivity gains turn out to be more related to capital
accumulation rather than to R&D expenditures. Supporting this argument, Castellani et al.
8
Since CIS considers only those firms which have more than ten employees and since smaller firms are likely
to be younger, our sample might be biased against the entrants. However, even if we were to consider all the
young firms, it is unlikely that the number of entrants would have increased substantially. Also, it would be
unlikely that among the entrants left out, a higher percentage of them would be engaging in formal R&D.

15
(2019) point out that complex and radical product innovation generally relies on formal R&D,
while process innovation is much more related to embodied technological change achieved by
investment in new machinery and equipment (see also Parisi et al. 2006). In Estonia, (i) 47%
of the firm-year observations from low-tech and medium-low-tech manufacturing and 23%
are from less knowledge intensive services, and (ii) process innovation being more prevalent
than product innovation (Table 1). This could potentially explain why productivity gains
from capital accumulation is (a) higher than that from R&D, and (b) higher for the Estonian
incumbents as compared to the German ones.
We find that the coefficient estimates of log(Employees) is negative for all firms. Now,
in the value-added production function in (4.3), the coefficient of log(Employees) for R&D
performing firms is βl+ = βl + βk + βr − 1 and βl0+ = βl0 + βk0 − 1 for non-R&D firms. The
negative estimates of βl+ and βl0+ , then suggest that the cumulative share of some of the
inputs is less than 1. In other words, it suggests that the returns to scale is decreasing.
As far as age is concerned, the estimates suggest that, within the two groups of firms, age
does not play a significant role as determinant of productivity. As a positive coefficient of
age is associated with learning effects which causes improvements in productivity with time
(Lubczyk and Peters 2020),9 the ACF estimates suggest that there are no learning effects.
Finally, we also find a robust evidence that firms located in northern Estonia, which is the
economic hub of the nation, are more productive; this suggests that firms in northern Estonia
benefit from a positive agglomeration effect.

5.2 Spillover Effects

Knowledge spillover effects, as we know, are highly relevant for both firms and policy makers,
as they point to situations where the benefits of accumulating knowledge in one firm result
in performance and productivity gains in a larger agglomeration or group of firms, further
resulting in a sub-optimal low level of R&D from a social point of view. Therefore, an
important question related to the productivity effects of R&D activities is the degree to
which a firm’s own R&D efforts impact other firms’ performances. Research on knowledge
spillovers at the micro-level finds that the different mechanisms through which spillovers occur
are indeed localised to a large extent (see Audretsch 1998; Storper and Venables 2004; Ponds
et al. 2010). Besides the importance of local labour markets and spin-off dynamics, studies
have emphasized the role of networking between individuals and between organizations as a
mechanism for knowledge spillovers that takes place at the regional level. For each firm, then,
we define external knowledge capital as the accumulation of knowledge that is generated by
other firms that are geographically close to the focal firm.
Since knowledge spillovers can take place both within and between industries, we further
differentiate the group of other firms into those firms that are within and those that are
outside of the focal firm’s own industry, and accordingly construct measures of inter- and
intra-industry knowledge capital (see section 4 for the definition of the two). These measures
of external knowledge capital are then added to the production functions in equations (4.3).
To test if there is differential impact of knowledge spillover for R&D performing firms and
9
A negative coefficient of firm age, at least for incumbents, indicates that these learning effects associated
with firm age become much smaller and phase out in later stages of firm life.

16
non-R&D firms, we interact the measures with two binary variables: D 0 and 1 − D 0 .
Now, in our data, we find that the average R&D expenditure of firms is higher in
counties/regions where there is a higher concentration (number of firms/geographical area
of the county) of firms. Moreover, we find that there is a positive correlation between R&D
expenses and the average R&D expenses of other firms situated in the same region. It, there-
fore, seems that location is not a random but a deliberate choice made by the firms. In other
words, the measures of external knowledge capital, through the endogenous choice of location,
and firm’s total factor productivity, ωjt , are correlated. The endogeneity of measures of ex-
ternal knowledge capital is accounted for by treating measures of external knowledge capital
and the dummy for north Estonia as state variables when using control function methods.
Table 3, columns 2a and 2b, illustrates the estimated results when measures of external
knowledge capital are added to the productivity functions. The estimates suggest that it
is mainly the incumbents who benefit from external knowledge produced outside the firms’
boundaries. However, among the incumbents, the R&D performing incumbents seem to gain
from only within industry external knowledge, whereas the incumbents who do not engage in
R&D activities gain from external knowledge both within and without their own industry.10
The estimates also suggest it is mainly the incumbents who benefit from knowledge capital
produced outside their own firm. In other words, these results indicate that mature firms are
on average better at incorporating diverse knowledge that stems from sources both within
and outside their industry. The results also show that for the incumbents, the social returns
to external knowledge capital are comparable to, if not higher than, the private returns to
R&D.
We find that non-R&D firms among the incumbents benefit more from external know-
ledge than innovating firms. Given that there exists complementarity between absorptive
capacity and external knowledge and that at sufficiently low levels of absorptive capacity
further increases in external knowledge may not increase marginal private incentives to build
absorptive capacity to benefit from it (see Aghion and Jaravel 2015), the results indicate
that incumbents that do not engage in R&D activities, nonetheless do invest in building
absorptivity capacity to benefit from external knowledge even though they do not invest in
STI based innovation. Second, since by definition non-R&D firms do not invest in ‘frontier
innovation,’ the results suggest that such incumbents are most likely benefiting from ‘techno-
logical adaptation or imitation,’ which is aided by geographical proximity to R&D intensive
firms operating in the same industry.

5.3 Productivity Implication of Technological Innovation for Entrants and

Incumbents
As mentioned earlier, about 30% of the firm-years have positive R&D expenses, but about
75% of the firm-years in the sample have innovated to introduce at least one new product
in the market or a new process. Also, about 7% of the firms that invested in R&D did not
innovate a new product or process. These suggest that more firms innovate through the DUI
10
Instead of the measure based on geographical distance, future research should consider estimating productiv-
ity response to measures of inter-industry external knowledge, which are based on ‘technological distance’ as
in Bloom et al. (2013) (see also Hall et al. 2010, on measuering spillovers).

17
mode of innovation rather than STI and that having positive R&D expenses is neither a
necessary nor a sufficient condition for innovation.
Now, it has been found that firms in the low- and medium-tech sectors are usually en-
gaged in the DUI mode of innovation while drawing on advanced science and technology
results available through the common knowledge bases (Robertson et al. 2009; Trott and
Simms 2017). Given that the majority of firms in our sample are in the low-tech and low-
medium-tech manufacturing and less knowledge intensive services (see Table 2), and many of
whom innovate without formally engaging in R&D, we, by estimating the average treatment
effect (ATE) of technological innovation, ∆, given in equation (4.5), assess how technological
innovations affect total factor productivity (TFP) and labour productivity consequently. As
D , P C )+ξ +ǫ ,
discussed in Appendix A, we first estimate the residual, ωjt +ǫjt = g(ωj,t−1 , Pjt jt jt jt
in equation (4.3) as a measure of TFP, and then use a matching method to estimate ∆. In
Table 7 we report the estimated ATEs.

[Table 6 about here]

[Table 7 about here]

To assess if there is complementarity between product and process innovation and to

understand how (a) only product innovation (b) only process innovation and (c) both product
and process innovation contribute to ∆, we also estimate the ATE of only product innovation
(∆10 ), the ATE of only process innovation (∆01 ), and the ATE of product and process
innovation (∆11 ) (see equation (4.7)). The control variables on the basis of which we match
innovators with the non-innovators for estimating the ATEs in (4.5) and (4.7) are stated
in the following. For estimating ∆ in (4.5), we match ‘exactly’ on year dummies (2006 to
2018) and 2-digit industry dummies, and match to ‘nearest neighbourhood’ on the lags of
the following variables: (1) residual, ωjt + ǫjt, as a proxy for TFP, (2) estimate of ωjt (see
footnote 11 in Appendix A), (3) product innovation, (4) process innovation (5) technological
innovation (6) log of R&D intensity (7) dummy for positive R&D, (8) labour productivity,
(9) log of number of employees, (10) log of capital stock, (11) log of material cost, (12) log of
intra-industry external knowledge, (13) log of inter-industry external knowledge, and (14) age.
However, because there were much fewer firms who only innovated products, only innovated
processes, or both, we could not match exactly on industry dummies while estimating ∆10 ,
∆01 , and ∆11 . We, therefore, matched exactly on year and to nearest neighbourhood on the
rest of the variables while estimating ∆10 , ∆01 , and ∆11 .
As we can see in Table 7, for productivity estimates from both the specifications, the
ATE, ∆, of technological innovation on productivity is larger for the entrants as compared to
the incumbents’. We find that the ATE ranges from 0.19 to 0.27 for the entrants, and same
for the incumbents ranges from 0.1 to 0.13. These ATEs, however, are in logarithmic scale;
at the mean, these difference in linear scale implies that an entrant who innovates is about
21% to 30% more productive than an entrant who does not. For the incumbents, it implies
that, on average, the incumbent firm that innovates is about 10% to 13% more productive
than the one which does not.
Given that a higher percentage of entrants (78%) compared to incumbents (72%) repor-
ted innovation, our results suggests that not only are entrants more innovative, but that

18
entrants are abler than incumbents in translating their innovations into productivity gains.
To understand how the two group of producers benefit from technological innovations, we
look at a more disaggregated picture by estimating the ATE of only product innovation (∆10 ),
the ATE of only process innovation (∆01 ), and the ATE of product and process innovation
(∆11 ). In Table 7, we can see that ∆10 is significantly positive for incumbents, implying that
incumbents who innovated new products only realized productivity gains. And with ∆01 sig-
nificantly positive for entrants, it implies that entrants who implemented innovative processes
only realized productivity gains. However, those firms that have innovated both product and
process realized the highest gain. The productivity of entrants who innovated both product
and process is about 35% higher than an average entrant’s, and such incumbents’ productivity
is about 13% higher than the average incumbent’s.
Since for the incumbents, the ATE of process innovation only did not turn out significant,
we can assume that for the incumbents ∆01 = 0; similarly we can assume that ∆10 is 0
for the entrants. We, therefore, find that for both the group of producers: ∆11 > ∆10 +
∆01 . In other words, we find the existence of complementarity between product and process
innovations. Now, as can be seen in Table 6, a higher percentage of innovating entrants (50%)
and incumbents (40%) innovate both products and processes. That is, both the groups of
producers, especially the entrants, prefer to invest in both the activities at the same time to
maximize their productivity.
Complementarity between product and process innovations arises when the returns to
implementing a product innovation are highest when the firm also implements a process
innovation. This occurs because product innovations, which shift out the demand curve,
induces the firm to produce a higher quantity. However, the returns from lowering the unit
cost through process innovations are highest precisely when quantity produced is highest, and
thus firms tend to implement product and process innovations together.
Athey and Schmutzler (1995) emphasize that flexibility in product designs and processes,
which lowers the adjustment costs of implementing product and process innovations, en-
hances the complementarity between product and process innovations. Since (a) entrants
are found to have gained from implementing process innovations and (b) complementarity
between product and process innovations is higher for the entrants, these results indicate that
entrants do leverage their flexible and non-hierarchical structures — due to which they lack
the structural inertia for reorganisation and which allow for faster decision-making processes
— for implementing the complementary process innovations, and consequently realize a higher
complementarity.
Before proceeding further, we would like to point out that since a smaller percentage of
firms — incumbents (31%) and entrants (29%) — reported engaging in formal R&D, this
result suggests that a lot of the productivity gains is being made through DUI mode of in-
novation. This assertion is also supported by our data: ‘DUI Exclusive mode of cooperation’
or cooperation with other enterprises for innovative activities excluding R&D was higher
(53%) for innovating firms that reported zero R&D, whereas ‘STI Exclusive mode of cooper-
ation’ was higher (73%) for firms with positive R&D. While our data indicates that many
innovations are likely a result of DUI mode of innovation, we are unable to identify which
innovations are a result of DUI mode of innovation and which are STI based. We are thus

19
unable to estimate the efficacy of STI based innovations vis-à-vis DUI based innovations,
or for that matter the efficacy of the combination of two, for the two group of producers.
We leave the exercise of assessing the efficacy of STI and DUI based innovations for future
research.
Finally, we compare the estimated TFP for the entrants and incumbents. As can be seen in
Table 7, on average, the TFP of the entrants is higher than that of the incumbents. This is in
contrast to labour productivity, which on average is found to be higher among the incumbents
than the entrants (see Table 1). Since labour productivity could change due to (a) changes
in technology, which improves the efficiency with which the factors operate, or (b) due to
changes in factor accumulation, it, therefore, seems that a higher labour productivity of the
incumbents is because of a higher capital-labour ratio than due to more efficient technology.
TFP, on the other hand, measures firm ability or efficiency, such as managerial talent, quality
of inputs, innovation, etc., that are not accounted for by the observed inputs (Syverson 2011).
That surviving young firms have above average productivity and that they grow faster than
their mature counterparts, has been documented elsewhere (Foster et al. 2008; Haltiwanger
et al. 2013).
Also, in Table 7 we can see that entrants’ TFP have higher standard deviation as compared
to the TFP for the incumbents. As discussed in Haltiwanger et al. (2013), young firms exhibit
an ‘up or out’ dynamic — they either grow fast on average or they exit. These ‘up or out’
dynamics imply that exiting young firms have very low productivity while surviving young
firms — where survival, more often than not, is because of successful innovation — have
above average productivity. Related to the ‘up or out’ dynamics, Foster et al. (2018) explain
that entry is associated ‘experimentation,’ which is critical for innovation and which results
high degree of within-industry productivity dispersion. Following this experimentation phase
is the ‘shakeout’ period, in which entrepreneurs that successfully innovate and/or adopt grow
while unsuccessful entrepreneurs contract and exit yielding productivity growth. Foster et al.
(2018) discuss a variety of mechanisms which can help understand how ‘experimentation’ can
generate heterogeneity in the factors that cause dispersion.

6 Concluding Remarks
A large proportion of productivity growth, a key driver of economic growth, is due to real-
location emanating from the entry and exit of firms. Since this reallocation is induced by
the selection effect due to the heterogeneous productivity impacts of R&D investment and
innovation undertaken by incumbents and entrants, in this paper we study the productivity
implications of (a) own R&D, (b) external knowledge, and (c) and technological innovations
for incumbents and entrants. The third exercise is motivated by the fact that, in our sample,
while only 30% of firm-year report investment in R&D, 74% report technological — product
and process — innovation. Studying the implications of R&D alone would have provided a
partial picture of the innovation activities of firms in Estonia and other CEE catching-up
economies. We, therefore, develop an empirical strategy to simultaneously estimate (1) the
productivity elasticities of R&D and external knowledge and (2) the impact of technological
innovations on productivity. Also, in studying the productivity implication of technological

20
innovation, we pay particular attention to complementarity between product and process
innovation for the two groups of producers. Besides, since physical capital stock embodies
technical change and innovation, we also study the productivity impact of capital accumula-
tion.
Several interesting conclusions emerge from our analysis. To start with, while both
entrants and incumbents gain significantly from investing in R&D, we do not find any dif-
ference in the average productivity elasticity with respect to own R&D for the two group
of producers. This is contrary to the findings in Acemoglu et al. (2018) and Lubczyk and
Peters (2020) for the technologically advanced economies, where entrants have been found to
be disproportionately abler in translating their R&D into productivity gains. Instead, we find
that the impact of innovation output — many of which are a result of DUI based innovation
— are similar to the results in the aforementioned papers: entrants who innovate are 21% to
30% more productive than entrants who do not; for the incumbents, the corresponding figures
are 10% to 13%. Furthermore, as we explored the sources of the productivity gain, we found
that the gain is largely due to complementarities between product and process innovation,
with entrants benefiting more from the complementarities than incumbents.
Estonian incumbents, on the other hand, are no less abler than German incumbents in
translating R&D or STI based innovative activities into productivity gains. This is despite
the adverse sectoral composition of the Estonian economy, which has a low share of sectors
that are R&D intensive and where returns to R&D are highest. Moreover, because the bulk
of STI based innovative activities in Estonia is carried out by the incumbents, they are the
primary carriers of STI based innovative activities.
We find that productivity elasticity with respect to capital is higher than with respect to
R&D, suggesting that in catching-up economies such as in Estonia, embodied technological
change through capital accumulation has been an important means, and likely more effective
than R&D, for generating productivity growth. However, since productivity elasticity with
respect to capital is higher for firms that report engaging in formal R&D, it suggests that
STI based innovative activities do help build absorptive capacity.
Finally, in results related to spillover effects, we find it is mostly the incumbents who
benefit from intra- and inter-industry knowledge that is generated by other firms — primarily
by the incumbents — situated in close proximity. However, the spillover effects are higher for
firms that do not reportedly engage in formal R&D, suggesting that these firms, too, work
on building absorptive capacity.
There are certain limitations in our paper, which are primarily due to lack of data.
First, as Community Innovation Surveys (CIS) consider firms with at least 10 employees,
the entrants included in these data may not be representative of the population of newly
born firms, which are unlikely to exceed this threshold in the first years of their existence.
Relying on CIS data could therefore limit the number of newly established firms. Second,
we consider only regional spillover effects; future studies could consider measures of external
knowledge that are based on technological distance between firms and/or forward and back-
ward linkages between industries.
Our results have important policy implications. The finding that technological innova-
tions benefit entrants disproportionately more than the incumbents points to the significant

21
impact innovative entry can have on economic dynamics. These results yet again emphasize
policies — in particular those that facilitate DUI mode of innovation — for promoting in-
novative entrepreneurship. However, our results also strongly suggest than in Estonia — a
catching-up economy — incumbents are the primary carriers of STI based innovative activ-
ities, which are instrumental in building absorptive capacity and help move economies closer
to the technological frontiers.
Our results then suggest that the policies advocated in Acemoglu et al. (2018) for techno-
logically advanced economies may not be applicable for a catching-up economy like Estonia.
Acemoglu et al. (2018) advocate that a large tax on the continued operation of incumbents
combined with a modest R&D subsidy to incumbent would be both growth and welfare in-
creasing. However, this policy only leverages the selection effect. As pointed out by Aghion
and Jaravel (2015, p. 539), the recommendations suggested by Acemoglu et al. (2018) do
not take into consideration the notion of absorptive capacity which could have “novel implic-
ations, as it creates another trade-off between incumbents (with a large stock of R&D and
a high absorptive capacity) and new firms.” Policy makers must, therefore, duly consider
the important role incumbents, as primary carriers of STI based innovative activities, play in
catching-up economies. As argued in Aghion and Jaravel (2015), because the social returns to
absorptive capacity are not fully internalised by the private market, private market forces may
fail to ensure the catching-up process; policies and institution that foster large incumbents
during the catching-up periods have historical antecedents (see Aghion et al. 2011).

References
Abadie, A. and Imbens, G. W. (2006). Large Sample Properties of Matching Estimators for Average
Treatment Effects. Econometrica, 74 (1), 235–267.

Acemoglu, D., Akcigit, U., Alp, H., Bloom, N. and Kerr, W. (2018). Innovation, Reallocation,
and Growth. American Economic Review, 108 (11), 3450–3491.

— and Cao, D. (2015). Innovation by Entrants and Incumbents. Journal of Economic Theory,
157 (C), 255–294.

Ackerberg, D. A., Caves, K. and Frazer, G. (2015). Identification Properties of Recent Produc-
tion Function Estimators. Econometrica, 83 (6), 2411–2451.

Aghion, P., Harmgart, H. and Weisshaar, N. (2011). Fostering Growth in CEE Countries: A
Country-tailored Approach to Growth Policy. In S. Radosevic and A. Kaderabkova (eds.), Chal-
lenges for European Innovation Policy, Chapters, 3, Edward Elgar Publishing.

— and Jaravel, X. (2015). Knowledge Spillovers, Innovation and Growth. The Economic Journal,
125 (583), 533–573.

Akcigit, U. and Ates, S. T. (2021). Ten Facts on Declining Business Dynamism and Lessons from
Endogenous Growth Theory. American Economic Journal: Macroeconomics, 13 (1), 257–298.

Asplund, M. and Nocke, V. (2006). Firm Turnover in Imperfectly Competitive Markets. The
Review of Economic Studies, 73 (2), 295–327.

Athey, S. and Schmutzler, A. (1995). Product and Process Flexibility in an Innovative Environ-
ment. The RAND Journal of Economics, 26 (4), 557–574.

22
Audretsch, B. (1998). Agglomeration and the Location of Innovative Activity. Oxford Review of
Economic Policy, 14 (2), 18–29.

Audretsch, D. (1995). Innovation and Industry Evolution. MIT Press.

Bartelsman, E., Haltiwanger, J. and Scarpetta, S. (2013). Cross-Country Differences in Pro-

ductivity: The Role of Allocation and Selection. American Economic Review, 103 (1), 305–34.

Baumann, J. and Kritikos, A. S. (2016). The link between R&D, Innovation and Productivity:
Are Micro Firms Different? Research Policy, 45 (6), 1263 – 1274.

Berchicci, L. and Tucci, C. (2009). Entrepreneurship, technology and schumpeterian innovation:

Entrants and incumbents. The Oxford Handbook of Entrepreneurship.

Bloom, N., Schankerman, M. and Van Reenen, J. (2013). Identifying Technology Spillovers and
Product Market Rivalry. Econometrica, 81 (4), 1347–1393.

Bruno, R. L., Douarin, E., Korosteleva, J. and Radosevic, S. (2019). Determinants of Pro-
ductivity Gap in the European Union: A Multilevel Perspective. IZA Discussion Papers 12542, Bonn.

Castellacci, F. and Natera, J. M. (2013). The dynamics of National Innovation systems: A Panel
Cointegration analysis of the Coevolution Between Innovative Capability and Absorptive Capacity.
Research Policy, 42 (3), 579–594.

Castellani, D., Piva, M., Schubert, T. and Vivarelli, M. (2019). R&D and Productivity in
the US and the EU: Sectoral Specificities and Differences in the Crisis. Technological Forecasting
and Social Change, 138, 279 – 291.

Cirillo, V., Martinelli, A., Nuvolari, A. and Tranchero, M. (2019). Only one way to skin
a cat? Heterogeneity and Equifinality in European National Innovation Systems. Research Policy,
48 (4), 905–922.

Coad, A., Segarra, A. and Teruel, M. (2016). Innovation and Firm Growth: Does Firm Age
Play a Role? Research Policy, 45 (2), 387–400.

Cohen, W. M. and Levinthal, D. A. (1989). Innovation and Learning: The Two Faces of R&D.
Economic Journal, 99 (397), 569–96.

Crépon, B., Duguet, E. and Mairesse, J. (1998). Research, Innovation And Productivity: An
Econometric Analysis At The Firm Level. Economics of Innovation and New Technology, 7 (2),
115–158.

Czarnitzki, D. and Delanote, J. (2012). Young Innovative Companies: The New High-Growth
Firms? Industrial and Corporate Change, 22 (5), 1315–1340.

Decker, R., Haltiwanger, J., Jarmin, R. and Miranda, J. (2014). The Role of Entrepreneurship
in US Job Creation and Economic Dynamism. Journal of Economic Perspectives, 28 (3), 3–24.

Doraszelski, U. and Jaumandreu, J. (2013). R&D and Productivity: Estimating Endogenous

Productivity. The Review of Economic Studies, 80 (4), 1338–1383.

European Commission (2011). Innovation Union Competitiveness Report 2011. Tech. rep.,
Directorate-General for Research and Innovation (European Commission).

Filippetti, A. and Peyrache, A. (2015). Labour Productivity and Technology Gap in European
Regions: A Conditional Frontier Approach. Regional Studies, 49 (4), 532–554.

23
Foster, L., Grim, C., Haltiwanger, J. and Wolf, Z. (2018). Innovation, Productivity Dispersion,
and Productivity Growth. Working papers, U.S. Census Bureau, Center for Economic Studies.

—, Haltiwanger, J. and Syverson, C. (2008). Reallocation, Firm Turnover, and Efficiency: Se-
lection on Productivity or Profitability? The American Economic Review, 98 (1), 394–425.

Garcia-Macia, D., Hsieh, C.-T. and Klenow, P. J. (2019). How Destructive Is Innovation?
Econometrica, 87 (5), 1507–1541.

Griffith, R., Redding, S. and Reenen, J. V. (2004). Mapping the Two Faces of R&D: Pro-
ductivity Growth in a Panel of OECD Industries. The Review of Economics and Statistics, 86 (4),
883–895.

Griliches, Z. (1979). Issues in Assessing the Contribution of Research and Development to Pro-
ductivity Growth. Bell Journal of Economics, 10 (1), 92–116.

Hall, B. (2011). Innovation and Productivity. Nordic Economic Policy Review, 17178.

—, Mairesse, J. and Mohnen, P. (2010). Measuring the Returns to R&D. In Handbook of the
Economics of Innovation, vol. 2, Chapter 24, Elsevier, pp. 1033–1082.

Haltiwanger, J., Jarmin, R. and Miranda, J. (2013). Who Creates Jobs? Small versus Large
versus Young. The Review of Economics and Statistics, 95 (2), 347–361.

Hopenhayn, H. A. (1992). Entry, Exit, and Firm Dynamics in Long Run Equilibrium. Econometrica,
60 (5), 1127–1150.

Huergo, E. and Jaumandreu, J. (2004). How does probability of innovation change with firm age?
Small Business Economics, 22, 193–207.

Jensen, M., Johnson, B., Lorenz, E. and Lundvall, B.-A. (2007). Forms of Knowledge and
Modes of Innovation. Research Policy, 36 (5), 680–693.

Jovanovic, B. (1982). Selection and the evolution of industry. Econometrica, 50 (3), 649–670.

Klepper, S. (1996). Entry, Exit, Growth, and Innovation over the Product Life Cycle. The American
Economic Review, 86 (3), 562–583.

Klette, T. J. and Kortum, S. (2004). Innovating Firms and Aggregate Innovation. Journal of
Political Economy, 112 (5), 986–1018.

Lacasa, I. D., Giebler, A. and Radošević, S. (2017). Technological Capabilities in Central and
Eastern Europe: An Analysis Based on Priority Patents. Scientometrics, 111 (1), 83–102.

Liik, M., Masso, J. and Ukrainski, K. (2014). The Contribution of R&D to Production Effi-
ciency in OECD Countries: Econometric Analysis of Industry-Level panel Data. Baltic Journal of
Economics, 14 (1-2), 78–100.

Lööf, H. and Johansson, B. (2014). R&D Strategy, Metropolitan Externalities and Productivity:
Evidence from Sweden. Industry and Innovation, 21 (2), 141–154.

Lubczyk, M. and Peters, B. (2020). Incumbents and Entrants as Carriers of Innovation and
Productivity Growth. Working papers, ZEW – Leibniz Centre for European Economic Research.

Mairesse, J. and Mohnen, P. (2010). Using Innovations Surveys for Econometric Analysis. Working
Paper 15857, National Bureau of Economic Research.

24
Masso, J., Eamets, R. and Philips, K. (2004). Creative Destruction and Transition: The Effects
of Firm Entry and Exit on Productivity Growth in Estonia. Iza discussion paper no. 1243, Institute
for the Study of Labour (IZA).

— and Vahter, P. (2008). Technological Innovation and Productivity in Late-Transition Estonia:

Econometric Evidence from Innovation Surveys. The European Journal of Development Research,
20 (2), 240–261.

Milgrom, P. and Roberts, J. (1990). The Economics of Modern Manufacturing: Technology,

Strategy, and Organization. American Economic Review, 80 (3), 511–28.

Olley, G. S. and Pakes, A. (1996). The Dynamics of Productivity in the Telecommunications

Equipment Industry. Econometrica, 64 (6), 1263–1297.

Ortega-Argilés, R., Piva, M. and Vivarelli, M. (2014). The Transatlantic Productivity Gap: Is
R&D the Main Culprit? Canadian Journal of Economics/Revue canadienne d’économique, 47 (4),
1342–1371.

—, — and Vivarelli, M. (2015). The Productivity Impact of R&D Investment: Are High-Tech
Sectors Still Ahead? Economics of Innovation and New Technology, 24 (3), 204–222.

Parisi, M. L., Schiantarelli, F. and Sembenelli, A. (2006). Productivity, Innovation and R&D:
Micro Evidence for Italy. European Economic Review, 50 (8), 2037–2061.

Peters, B. (2009). Persistence of innovation: Stylised Facts and Panel Data Evidence. The Journal
of Technology Transfer, 34 (2), 226–243.

—, Roberts, M. J., Vuong, V. A. and Fryges, H. (2017). Estimating dynamic R&D choice: an
analysis of costs and long-run benefits. The RAND Journal of Economics, 48 (2), 409–437.

Ponds, R., van Oort, F. and Frenken, K. (2010). Innovation, Spillovers and University–Industry
Collaboration: An Extended Knowledge Production Function Approach. Journal of Economic Geo-
graphy, 10 (2), 231–255.

Radosevic, S. (1998). The Transformation of National Systems of Innovation in Eastern Europe:

Between Restructuring and Erosion. Industrial and Corporate Change, 7 (1), 77–108.

— and Yoruk, E. (2018). Technology Upgrading of Middle Income Economies: A New Approach
and Results. Technological Forecasting and Social Change, 129 (C), 56–75.

Raymond, W., Mairesse, J., Mohnen, P. and Palm, F. (2015). Dynamic models of r&d, innova-
tion and productivity: Panel data evidence for dutch and french manufacturing. European Economic
Review, 78, 285 – 306.

—, Mohnen, P., Palm, F. and van der Loeff, S. S. (2010). Persistence of Innovation in Dutch
Manufacturing: Is It Spurious? The Review of Economics and Statistics, 92 (3), 495–504.

Robertson, P., Smith, K. and von Tunzelmann, N. (2009). Innovation in low- and medium-
technology industries. Research Policy, 38 (3), 441–446.

Rovigatti, G. and Mollisi, V. (2018). Theory and Practice of Total-Factor Productivity Estima-
tion: The Control Function Approach using Stata. The Stata Journal, 18 (3), 618–662.

Salinger, M. and Summers, L. H. (1983). Tax Reform and Corporate Investment: A Microecono-
metric Simulation Study, University of Chicago Press, pp. 247–288.

25
Savignac, F. (2008). The Impact of Financial Constraints on Innovation: What can be learned from
a Direct Measure? Economics of Innovation and New Technology, 17 (6), 553–569.

Storper, M. and Venables, A. J. (2004). Buzz: Face-to-Face Contact and the Urban Economy.
Journal of Economic Geography, 4 (4), 351–370.

Syverson, C. (2011). What Determines Productivity? Journal of Economic Literature, 49 (2), 326–
65.

Trott, P. and Simms, C. (2017). An Examination of Product Innovation in Low-and Medium-

Technology Industries: Cases from the UK Packaged Food Sector. Research Policy, 46 (3), 605–623.

26
Appendices

A Identification of Average Treatment Effect of Innovation

and Testing for Complementarity Between Product and In-
novation
Substituting the specification for ωjt given in equation (4.4) in the log-linear production
function in equation (4.3), we obtain

0
yjt =(1 − Djt )(β + βl+ ljt + βk kjt + βr rjt + βe ejt ) + Djt
0
(β 0 + βl0+ ljt + βk0 kjt + βe0 ejt )
D C
+ g(ωj,t−1 , Pjt , Pjt ) + ξjt + ǫjt . (A.1)

D , P C ), as discussed in section 4, is estim-

The coefficients, β’s, and the function, g(ωj,t−1 , Pjt jt
ated using the control function methodology by Ackerberg et al. (2015) (ACF). For notational
convenience from here on, we drop the firm subscript, j.
In this appendix, we discuss how we (i) estimate the average treatment effect (ATE) of
the various kinds of innovation (∆ in Eq. (4.5) and ∆10 , ∆01 , ∆11 in (4.7) ) and (ii) assess
for complementarity between product and process innovation.
Now, as discussed in the main text, the function, g(ωt−1 , PtD , PtC ), in equation (A.1) is
estimated non-parametrically by approximating it by a polynomial function of the arguments
of g(ωt−1 , PtD , PtC ). So, if, for example, ωt−1 is calculated as in footnote 11, one could em-
ploy the estimated coefficients of the polynomial to obtain estimates of g1∨1 (ωt−1 , .) and
g0∧0 (ωt−1 , .), where the two are defined in equation (4.6), for every ωt−1 , and then by aver-
aging over ωt−1 one can obtain, for example, ∆, the average treatment effect of technological
innovation:

∆ = E[g1∨1 (ωt−1 , .)] − E[g0∧0 (ωt−1 , .)]. (A.2)

For inference, we could either (a) first obtain bootstrapped standard errors of the coefficients
of the polynomial function, and then use the delta method to obtain the standard error of
∆, or (b) directly bootstrap ∆. Both procedures are, however, computationally intensive.
Rather than following these laborious procedures of computing the ATEs and their standard
errors, we, as discussed below, develop a simple procedure to estimate the ATEs.
In the following we show that we can still obtain ∆ in (A.2) without using the estimates
of the function, g(ωt−1 , .). The alternative procedure entails first estimating the residuals

rt ≡ yt − (1 − Dt0 )(β + βl+ lt + βk kt + βr rt + βe et ) + Dt0 (β 0 + βl0+ lt + βk0 kt + βe0 et )

= g(ωt−1 , PtD , PtC ) + ξt + ǫt

in equation (A.1) and then using the matching methods for estimating the ATE, ∆. The
alternative procedure works because these residuals, which are a proxy for the total factor
productivity (TFP), contain information on g(ωt−1 , .). The residuals for firm-years that
innovated are rt (It = 1) = g1∨1 (ωt−1 , .) + ξt + ǫt , and for those that did not are rt (It = 0) =

27
g0∧0 (ωt−1 , .) + ξt + ǫt .
Having obtained the residuals, g(ωt−1 , .) + ξt + ǫt , to estimate the impact of technological
innovations on productivity, we can obtain the difference

E[g1∨1 (ωt−1 , .) + ξt + ǫt ] − E[g0∧0 (ωt−1 , .) + ξt + ǫt ], (A.3)

where the expectation is taken over (ωt−1 , ξt , ǫt ).

Note that we only know rt (It = 1) for firm-years that have innovated and rt (It = 0) for
firms that did not innovate. Our problem is, therefore, same as the missing data problem
in treatment effect literature: rt (It = 0) is unobserved for firm-years that innovated, and
rt (It = 1) is unobserved for firm-years that did not innovate. The comparison, therefore,
between E[g1∨1 (ωt−1 , .) + ξt + ǫt ] and E[g0∧0 (ωt−1 , .) + ξt + ǫt ] for assessing the effect of
innovation can only be possible if ωt−1 for firms that have innovated (treatment group) are
same as ωt−1 for firms that did not (control group).
To select firms in control group with ωt−1 similar to the firms in the treatment group, we
resort to the matching method in Abadie and Imbens (2006). The method rigorously selects
control units with Xt−1 similar to the treated units. The control variables (or ‘observable
confounders’), Xt−1 , on the basis of which we match firms in treatment group to the firms
in the control group are variables that likely affect both the treatment (innovation) and the
outcome (TFP). In Xt−1 we include (i) ωt−1 ,11 (ii) all the state variables and control variables
D and P C . The control variables, X
from period t − 1, and (iii) Pt−1 t−1 t−1 , are mean independent
of the error terms, ξt + ǫt (see also discussion following Eq. (4.4)).
The identification assumption is that conditional on Xt−1 , the treatment is independent
of the ‘potential’ outcome, rt (It ). In the other words, for the matched sample, the treatment
(here innovation) can then be considered to be randomly assigned between the treatment and
the control group. Under this assumption, for the matched sample, we, therefore, have

∆(Xt−1 ) = E(rt (It = 1) − rt (It = 0)|Xt−1 )

= E[g(ωt−1 , .) + ξt + ǫt |It = 1, Xt−1 ] − E[g(ωt−1 , .) + ξt + ǫt |It = 0, Xt−1 ]
= E[g1∨1 (ωt−1 , .) + ξt + ǫt |Xt−1 ] − E[g0∧0 (ωt−1 , .) + ξt + ǫt |Xt−1 ]
= E[g1∨1 (ωt−1 , .)|Xt−1 ] − E[g0∧0 (ωt−1 , .)|Xt−1 ], (A.4)

where the last equality follows also because ξt + ǫt is mean independent of Xt−1 .
Under the overlap assumption, Pr(It = 1|Xt−1 ) < 1, (see Abadie and Imbens 2006)
11
In ACF, firm’s intermediate input demand is given by mt = ft (lt , kt , rt , et , xt , ωt ), where xt is the set of
additional state variables, and the function, mt = ft (., ωt ), is strictly monotonic in the unobserved TFP, ωt .
Given the monotonicity assumption, a proxy for ωt is obtain by inverting the intermediate input demand
function: ωt = ft−1 (lt , kt , rt , et , xt , mt ). STATA’s prodest command gives an estimate of ωt as

φ̂(Dt0 , lt , kt , rt , et , xt , mt ) − [(1 − Dt0 )(β̂ + β̂l+ lt + β̂k kt + β̂r rt + β̂e et ) + Dt0 (β̂ 0 + β̂l0+ ljt + β̂k0 kt + β̂e0 et )],

where φ(.), which is estimated in the first stage, is given by

φ(Dt0 , lt , kt , rt , et , xt , mt ) =[(1 − Dt0 )(β + βl+ lt + βk kt + βr rt + βe et ) + Dt0 (β 0 + βl0+ lt + βk0 kt + β̂e0 et )]

+ ft−1 (lt , kt , rt , et , xt , mt ).

Note that given ωt = ft−1 (lt , kt , rt , et , xt , mt ), the next period productivity is ωt+1 = g(ωt , PtD , PtC ) + ξt , and
it is g(ωt , PtD , PtC ) that is our quantity of interest, not ωt = ft−1 (lt , kt , rt , et , xt , mt ).

28
the difference on the right-hand side of (A.4) is identified for almost all Xt−1 in the sup-
port of Xt−1 . Therefore, the average effect of the treatment can be recovered by averaging
E[g1∨1 (ωt−1 , .)|Xt−1 ] − E[g0∧0 (ωt−1 , .)|Xt−1 ] over the distribution of Xt−1 . That is,

E[∆(Xt−1 )] = E[g1∨1 (ωt−1 , .)] − E[g0∧0 (ωt−1 , .)] = ∆, (A.5)

which is what we wanted to estimate in (A.2).

We now come to assessing of complementarity between product, PtD = 1, and process,
PtC = 1 innovation. The study of complementarities between activities can be traced back
to the theory of supermodularity (Milgrom and Roberts 1990). According to this theory,
the necessary condition for product and process innovation to be complementary is that the
expectation, E[g(ωt−1 , PtD , PtC )], be supermodular in PtD and PtC :

E[g1∧1 (.)] − E[g0∧1 (.)] ≥ E[g1∧0 (.)] − E[g0∧0 (.)], (A.6)

where

D C D C
g1∧0 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 1 and Pjt = 0,
D C D C
g0∧1 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 0 and Pjt = 1,
D C D C
g1∧1 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 1 and Pjt = 1 and
D C D C
g0∧0 (.) ≡ g(ωj,t−1 , Pjt , Pjt ) such that Pjt = 0 and Pjt = 0.

The inequality in (A.6) can be interpreted as follows: on average, implementing a product

innovation when the firm also implements a process innovation has a higher incremental effect
on productivity than implementing a product innovation in isolation.
We can write the inequality in (A.6) as

E[g1∧1 (.)] − E[g0∧0 (.)] ≥(E[g0∧1 (.)] − E[g0∧0 (.)]) +(E[g1∧0 (.)] − E[g0∧0 (.)])
⇐⇒ ∆11 ≥ ∆01 + ∆10 , (A.7)

where ∆10 is the ATE of only product innovation, ∆01 is the ATE of only process innovation,
and ∆11 , the ATE of both product and process innovation. That is, if the productivity
impact of product and process innovation exceeds the sum of the impacts of only product
innovation and only process innovation, then the product and process innovation can be said
to be complementary. The ATEs, ∆10 , ∆01 , and ∆11 , can be estimated using the matching
method as in (A.5).

29
 Table 1: Test of Equality of Means between Entrants and Incumbents

Entrants Incumbents Difference Pr(|T | > |t|)

log(Value-Added Productivity ) 10.2 10.33 -0.13 0.00
log(Gross Output Productivity) 11.16 11.27 -0.11 0.00
log(No. of Employees) 3.5 3.75 -0.25 0.00
log(Capital) 12.4 13.36 -0.96 0.00
log(Material Cost) 13.57 14.01 -0.44 0.00
log(Investment in Fixed Assets) 11.17 11.53 -0.36 0.00
log(R&D Expenditure) 7.17 7.78 -0.61 0.000
Dummy for Positive R&D Expenditure 0.29 0.31 -0.02 0.16
Dummy for Innovator 0.78 0.72 0.06 0.00
Dummy for Product Innovation 0.56 0.45 0.11 0.00
Dummy for Process Innovation 0.61 0.56 0.04 0.00
Dummy for North 0.59 0.54 0.05 0.00
log(Intra-Industry R&D) 4.27 4.56 -0.28 0.02
log(Inter-Industry R&D) 9.39 10.07 -0.68 0.00
Note: (a) The two sample t-test assumed unequal variances. (b) The data used for the
tests is the estimation sample.

Table 2: Sectoral Distribution of Estonian Firms in the Estimation Sample

Percentage Positive R&D Technological  R&D 

log
of Firms Innovation Employees
Entrants
High-Tech Manufacturing 2.3 37.5 97 3.41
Medium High-Tech Manufacturing 10.14 35.5 86 2.54
Medium Low-Tech Manufacturing 22.01 22.9 75 1.57
Low-Tech Manufacturing 28.42 27.1 81 1.84
Knowledge Intensive Services 19.93 43 82 3.47
Less Knowledge Intensive Services 17.19 19.2 69 1.15
Total 100 28.9 78 2.1
Incumbents
High-Tech Manufacturing 2.26 54.1 83 4.7
Medium High-Tech Manufacturing 11.02 46 82 3.78
Medium Low-Tech Manufacturing 15.13 27.1 71 2
Low-Tech Manufacturing 31.62 28.2 77 1.94
Knowledge Intensive Services 15.59 41 72 3.52
Less Knowledge Intensive Services 24.39 21.5 64 1.68
Total 100 30.7 72 2.4
No. of Entrants: 1500, No. of Incumbents: 6086.
Note: (a) All figures excepts for log(R&D/Employees) are in percentages. (b) Technolo-
gical classification borrowed from Eurostat is based on NACE Revision 2 2-digit level.

30
Table 3: Productivity Effects of R&D and Capital Accumulation for Entrants and
Incumbents

Specification: Baseline
Specification: Baseline +
External Knowledge
Incumbent Entrant Incumbent Entrant
(1a) (1b) (2a) (2b)
log(Employees)×(1 − D ) 0 -0.16 ∗∗∗ -0.035 -0.234 ∗∗∗ -0.188∗∗∗
(0.001) (0.011) (0.001) (0.013)
log(Employees)×D 0 -0.147∗∗∗ -0.109∗∗∗ -0.141∗∗∗ -0.101∗∗∗
(0.001) (0.03) (0.001) (0.007)
log(Capital/Employees) ×(1 − D 0 ) 0.304∗∗∗ 0.194∗∗∗ 0.300∗∗∗ 0.186∗∗∗
(0.002) (0.023) (0.001) (0.009)
log(Capital/Employees) ×D 0 0.293∗∗∗ 0.198∗∗∗ 0.289∗∗∗ 0.174∗∗∗
(0.002) (0.016) (0.002) (0.007)
log(R&D/Employees) 0.046 ∗∗∗ 0.064 ∗∗∗ 0.081 ∗∗∗ 0.079∗∗∗
(0.005) (0.02) (0.002) (0.009)
0
D : Dummy Zero R&D 0.047 ∗∗∗ 0.067 ∗∗∗ -0.308 ∗∗∗ -0.276∗∗∗
(0.001) (0.009) (0.001) (0.013)
log(Intra-industry R&D)×(1 − D 0 ) 0.022∗∗∗ 0.019∗∗∗
(0.003) (0.001)
log(Intra-industry R&D)×D 0 0.025∗∗∗ 0.01
(0.002) (0.001)
log(Inter-industry R&D)×(1 − D 0 ) -0.009 -0.04
(0.007) (0.009)
log(Inter-industry R&D) ×D 0 0.035∗∗∗ 0.015∗∗
(0.001) (0.001)
Age 0.000 0.017 0.003 -0.005
(0.001) (0.019) (0.002) (0.008)
North Estonia 0.326∗∗∗ 0.268∗∗∗ 0.192∗∗∗ 0.314∗∗∗
(0.002) (0.009) (0.000) (0.009)
Test of Equality of Elasticity of Productivity with respect to Capital for firms with and
without R&D.
Test Statistic: C 6.92∗∗∗ 0.06 6.03∗∗ 0.78
2
Pr(χ (1) > C) 0.008 0.799 0.014 0.376
No. of Observations 6086 1500 5877 1464
No. of Firms 2443 1072 2394 1052
Note: Dependent Variable: log(Value-Added/Employees). Every specification includes
Time and Industry Dummies.
Significance levels : ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%

31
 Table 4: Comparison of Elasticity of Productivity with respect to Own R&D and Capital with those obtained for Germany

Specification: Baseline Specification: Baseline + External Knowledge

Method: ACF Method: OP Method: ACF Method: OP
Incumbent Entrant Incumbent Entrant Incumbent Entrant Incumbent Entrant
(1a) (1b) (1c) (1d) (2a) (2b) (2c) (2d)
Estonia
log(Capital/Employees) ×(1 − D 0 ) 0.304∗∗∗ 0.194∗∗∗ 0.179∗∗∗ 0.164∗∗∗ 0.300∗∗∗ 0.186∗∗∗ 0.193∗∗∗ 0.128∗∗∗
(0.002) (0.023) (0.017) (0.057) (0.001) (0.009) (0.002) (0.033)
log(Capital/Employees) ×D 0 0.293∗∗∗ 0.198∗∗∗ 0.148∗∗∗ 0.116∗∗ 0.289∗∗∗ 0.174∗∗∗ 0.176∗∗∗ 0.074∗∗∗
(0.002) (0.016) (0.017) (0.051) (0.002) (0.007) (0.002) (0.027)
log(R&D/Employees) 0.046 ∗∗∗ 0.064 ∗∗∗ 0.026 ∗∗ 0.032∗∗ 0.081∗∗∗ 0.079∗∗∗ 0.062∗∗∗ 0.037∗∗
32

(0.005) (0.02) (0.012) (0.015) (0.002) (0.009) (0.002) (0.009)

Germany
log(Capital/Employees) 0.127∗∗∗ 0.281∗∗∗0.076∗∗∗ 0.147∗∗∗ 0.133∗∗∗ 0.255∗∗∗ 0.021 0.148∗∗∗
(0.000) (0.001) (0.026) (0.029) (0.002) (0.003) (0.026) (0.007)
log(R&D/Employees) 0.057∗∗∗ 0.125∗∗∗ 0.024∗∗∗ 0.029∗∗∗ 0.050∗∗∗ 0.120∗∗∗ 0.022∗∗∗ 0.026∗∗∗
(0.002) (0.002) (0.004) (0.010) (0.002) (0.003) (0.002) (0.009)
Note: Lubczyk and Peters (2020) estimated the elasticities using two different control function methods: one, the method by ACF, and the
other by Olley and Pakes (1996) (OP). For the purpose of comparison, we also estimate using both the methods. The dependent variable
is log(Value-Added/Employees) when employing ACF and log(Revenue/Employees) when employing OP. Also, though not reported here,
log(Material/Employees) interacted with D 0 and 1 − D 0 are additional explanatory variables when using the OP method, where D 0 is a
dummy variable that takes value 1 if the firm reports zero R&D expense. Every specification includes Time and Industry Dummies.
Significance levels : ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%
Table 5: Productivity Elasticity with respect to R&D for (a) firms in High-Tech
and Medium-High-Tech Manufacturing and (b) firms in the Remaining Sectors

Specification: Baseline
Specification: Baseline +
External Knowledge
Incumbent Entrant Incumbent Entrant
log(R&D/Employees) ×DHM 0.061 ∗∗∗ 0.045 ∗∗ 0.094 ∗∗∗ 0.074∗∗∗
(0.002) (0.021) (0.006) (0.005)
log(R&D/Employees) ×(1 − DHM ) 0.037 ∗∗∗ 0.080 ∗∗∗ 0.085 ∗∗∗ 0.091∗∗∗
(0.001) (0.022) (0.001) (0.005)
Test of Equality of Elasticity of Productivity with respect to R&D for (a) firms in High-Tech
and Medium-High-Tech Manufacturing and (b) firms in the Remaining Sectors
Test Statistic: C 10.04∗∗∗ 1.51 68.4∗∗∗ 4.50∗∗
Pr(χ2 (1) > C) (0.00) (0.22) (0.00) (0.03)
Note: DHM is binary variable that takes value 1 if the firm is in the high-tech manufacturing
or medium-high-tech manufacturing and 0 otherwise. For lack of space, we only display the
productivity elasticities with respect to R&D for (a) firms in the high-tech and medium-high-
tech manufacturing and (b) firms in the remaining sectors.
Significance levels : ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%

D = 1) and Process (P C = 1)
Table 6: Description of innovative activity: Product (Pjt jt
Innovation

Incumbents Entrants
C
Pjt C =0
= 1 Pjt Total C
Pjt = 1 PjtC =0 Total
D =1
Pjt 1695 1650 3345 D =1
Pjt 329 333 662
(27.85) (27.11) (54.96) (21.93) (22.2) (44.13)
D =0
Pjt 956 1785 2741 D =0
Pjt 260 578 838
(15.71) (29.33) (45.04) (17.33) (38.53) (55.87)
Total 2651 3435 6086 Total 589 911 1500
(43.56) (56.44) (100) (39.27) (60.73) (100)

Note: Cell percentage in parentheses.

33
Table 7: Average Treatment Effect of Technological Innovation on Estimated Total
Factor Productivity for Entrants and Incumbents

Specification: BaselineSpecification: Baseline +

External Knowledge
Incumbent Entrant Incumbent Entrant
ATE of Product and/or Process 0.129 ∗∗ 0.267 ∗∗∗ 0.102 ∗ 0.191∗∗∗
Innovation, Ijt, on Productivity: ∆ (0.051) (0.072) (0.054) (0.079)
No. of matches for computing the ATE 1001 378 1000 346
ATE of only Product 0.097∗ 0.193 0.116∗∗ 0.135
Innovation on Productivity: ∆10 (0.057) (0.123) (0.058) (0.196)
No. of matches for computing the ATE 934 123 910 114
ATE of only Process 0.048 0.227∗∗ 0.066 0.223∗∗
Innovation on Productivity: ∆01 (0.046) (0.089) (0.047) (0.107)
No. of matches for computing the ATE 1286 172 1250 158
ATE of Product and Process 0.134 ∗∗∗ 0.305 ∗∗∗ 0.122 ∗∗∗ 0.300∗∗∗
Innovation on Productivity: ∆11 (0.043) (0.093) (0.045) (0.104)
No. of matches for computing the ATE 1414 198 1381 183
Mean of Productivity 7.496 8.235 7.879 8.894
Standard Deviation of Productivity 0.742 0.832 0.735 0.815
Note: While estimating estimating ∆, we matched ‘exactly’ on year dummies (2006 to 2018) and
2-digit industry dummies, and matched to ‘nearest neighbourhood’ on the lags of the variables
mentioned in the text. However, because there were much fewer firms who only innovated
products, only innovated processes, or both, we could not match ‘exactly’ on industry dummies
while estimating ∆10 , ∆01 , and ∆11 . As a result of two different matching criteria, we find that
for the incumbents, the number of matches are higher when estimating ∆10 , ∆01 , and ∆11 even
as the number of incumbents who only innovated products, or only innovated processes, or both
were, by definition, smaller than the number of incumbents who innovated either products, or
processes, or both.
Significance levels : ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%

34