Are science cities fostering firm innovation? Evidence from Russian regions
1. Introduction Background Data Methodology Results Conclusion
Are science cities fostering innovation?
Evidence from Russia
Helena Schweiger1 Alexander Stepanov1 Paolo Zacchia2
1EBRD
2IMT Lucca
SITE Academic Conference: 25 years of transition, Stockholm,
5 December 2016
2. Introduction Background Data Methodology Results Conclusion
Motivation
Innovation is a key driver of economic growth
Innovation tends to be spatially clustered (spillovers)
If innovation is an externality, what role for the government?
Either indirect (incentives) or direct (investment)
Both approaches can be place-based. A classical example is
the Silicon Valley, tracing roots in U.S. military investment
Their effect is difficult to evaluate
In Russia, a debate of particular relevance:
Innovation is essential to diversify the economy
Russia possesses excellent human capital resources as well as a
tradition of localized R&D policies: science cities
Assessing their effect is crucial to formulate growth policies
3. Introduction Background Data Methodology Results Conclusion
Research question
Question
Do innovation-focused place-based policies have an impact on local
development? What is their effect on innovation and productivity,
both at the municipal and firm level?
Contribution
1 First paper to evaluate the legacy of“innovation enclaves”in
the former Soviet Union on innovation in present-day Russia
2 We assess the impact on science cities both at the municipal
and at the firm level, employing two unique datasets
3 Municipal level data: a combination of geographical, historical
and present characteristics of Russian municipalities
4 Firm-level data from BEEPS V: new and accurate measures of
product and process innovation
4. Introduction Background Data Methodology Results Conclusion
Preview of the results
Methodology
We employ a variety of matching methods to correct for:
the selection bias due to the government’s choice of location;
the self-selection of firms into science cities.
Main Results
1 Today’s science cities are responsible for larger patent
production than otherwise similar“ordinary”municipalities
2 No evidence that, all else equal, firms in science cities are
different from firms in otherwise similar“ordinary”
municipalities
5. Introduction Background Data Methodology Results Conclusion
Related literature
1 (Localised) knowledge spillovers:
Jaffe et al. (1993), Moretti (2004, 2011), Bloom, Schankerman
and Van Reenen (2013), Lychagin et al. (forthcoming)
2 Evaluation of place-based policies:
Short-run: Neumark and Kolko (2010), Ham et al. (2011),
Albouy (2012), Busso et al. (2013), Wang (2013)
Long-run: Kline and Moretti (2014), Ivanov (2016)
3 Knowledge-focused place-based policies:
Felsenstein (1994), Westhead (1997), Siegel et al. (2003),
Yang et al. (2009), Falck et al. (2010)
4 Military and R&D:
Moretti, Steinwander and Van Reenen (2015)
6. Introduction Background Data Methodology Results Conclusion
Innovation system in the Soviet Union
Best resources were allocated to sectors considered vital for
national security, military (2/3 of R&D spending)
Model: special-regime enclaves aimed at fostering innovation
Main research areas, in order of relevance:
Aviation, rocket and space science
Nuclear physics
Electronics, mechanics
Chemistry and chemical physics
Biology and biochemistry
Civilian applications of technological advances were limited
(example: lag in personal computer development in late years)
7. Introduction Background Data Methodology Results Conclusion
Innovation system in the Soviet Union (cont.)
1 Early 1930s: Experimental Design Bureaus (sharashki), part
of the Soviet Gulag labour camp system
2 From mid-1930s: Science cities – localities with high
concentration of R&D facilities
About 2/3 established in existing cities/settlements, the rest
built from scratch
Those whose main objective was to develop nuclear weapons,
missile technology, aircraft and electronics, were closed,
represented only on classified maps, to maintain security and
privacy
Academic towns, to foster development of Siberia
9. Introduction Background Data Methodology Results Conclusion
Innovation system in Soviet Union/Russia
Soviet Union
R&D spending as % of GDP considerable, remarkable successes
were achieved but: no competition, lack of quality resources in the
civil economy, and bias towards large-scale projects, non-existent
small enterprise sector
Russia
Public funding dried up with the collapse of the Soviet Union.
Today: R&D spending much lower, almost 3/4 of R&D conducted
by public organisations, funded mainly from the federal budget
State programs designed to support economic development and
modernization via technological innovation and commercialization of
innovation since the early 1990s.
Science cities, innovation technology centres, technology parks,
incubators, Special Economic Zones, Skolkovo
10. Introduction Background Data Methodology Results Conclusion
Three unique, interconnected datasets
1 Database with information on science cities, based on
Aguirrechu (2009) and publicly available information
2 Firm-level data at the plant level from the BEEPS V survey,
including the novel innovation module.
37 Russian regions, 4220 face-to-face interviews conducted
between August 2011 and October 2012
Additional information allows more accurate measurement of
product and process innovation
3 Municipal-level data of Soviet/Russian municipal units (raion
level), containing geographical, climate, population,
education, innovation and economy-related data, both historic
and present-day
11. Introduction Background Data Methodology Results Conclusion
Municipal-level database
1 Geographical data: Location of coastline, lakes and rivers,
railroads in 1943, average monthly temperatures 1960-1990,
area of the municipality, GPS coordinates of the centre of
municipality and various distances calculated in QGIS
2 Data on factories, research and design establishments of the
Soviet defence industry from Dexter and Rodionov (2016)
3 Population data from the first post-World War II census
conducted in January 1959
4 Number of higher education institutions in 1959 from De Witt
(1961) and number of R&D institutes in 1959 from various
open sources (primarily Wikipedia)
5 Population with higher education or PhD/doctoral degrees
from 2010 Russian census
6 Patent data from WIPO/EPO/USPTO geolocated patents,
2006-2015
7 Nighttime lights data from NOAA, 1992-1994 and 2009-2011
12. Introduction Background Data Methodology Results Conclusion
Methodology: Summary
Science cities and non-science cities are different in their
observable characteristics
However: selection criteria of the Soviet government were
arguably not very much influenced by market forces or internal
political considerations
Less potential for bias(es) caused by unobservable factors
To account for differences in the observables, we employ
matching techniques:
1 For the municipal analysis, we use stratification and nearest
neighbour propensity score matching
2 For the firm-level analysis, we argue that coarsened exact
matching (CEM) is preferable due to two-way selection
13. Introduction Background Data Methodology Results Conclusion
Matching municipalities
Ideal quasi-experiment for the municipal analysis: compare
places that were chosen to become science cities (Dc = 1)
against those that were marginally discarded (Dc = 0)
In this case: ATT = long-run effect
We match municipalities on a mix of historical and geographic
variables using two alternative methods of propensity score
matching:
Stratification matching
Nearest neighbour matching
14. Introduction Background Data Methodology Results Conclusion
Matching municipalities (cont.)
Set of observable characteristics Xc for matching:
Geographical/climate characteristics: average temperature in
January and July; altitude; longitude, latitude; access to and
distance from lake, river and coast; distance from border; area
of the municipality
Historic characteristics: population in 1959; access to and
distance from railroad in 1943; number of higher education
institutions in 1959; number of scientific R&D institutions in
1947; number of defence plants (except R&D) in 1947
Main outcomes of interest Yc:
Total and fractional patents
Nighttime lights, 2009-2011
Total population, population with higher education and
population with PhD and doctoral degrees from the 2010
Russian Census
15. Introduction Background Data Methodology Results Conclusion
Matching firms in municipalities
Firm-level analysis: a two-way (dual) selection problem
Firm location choice is endogenous. Preference for science
city location Dic changes with firm characteristics Zic
But firms also choose over other municipal characteristics Xc,
which co-vary with Dic
Thus: simply matching over Zic is biased. Group differences
may be due to differences in Xc (e.g. population, education)
rather than science city status Dic
Further complications: a. Zic poorly predicts Dic, while b. Xc
identically predicts it for firms in the same city c
Common support issues, propensity score matching techniques
underperform
16. Introduction Background Data Methodology Results Conclusion
Matching firms in municipalities (cont.)
We apply CEM over both Xc and Zic: similar firms in similar
cities
Iacus, King and Porro (2009, 2011, 2015), King and Nielsen
(2016) argue for CEM against propensity score matching
In particular, propensity score matching weak when the
propensity score model has little power
Covariates employed for CEM:
Firm-level characteristics Zic : Size, age, foreign and/or state
controlled, market scope (local, national, international), % of
university educated employees, manufacturing/retail dummy
Municipal-level characteristics Xic : Population in 2010,
Postgraduate population in 2010, Fractional patents
We force exact matching within administrative divisions
Outcomes Wic: 1. innovation and 2. productivity measures
1 BEEPS innovation module: product and/or process innovation
dummies, R&D dummy, innovation-dependent share of sales.
2 Solow residual / TFP, raw labor productivity
17. Introduction Background Data Methodology Results Conclusion
Matching municipalities: Common support
(pre-matching)
19. Introduction Background Data Methodology Results Conclusion
Municipal-level results: Stratification matching
(1) (2) (3) (4) (5) (6)
All Frac. Night Population Graduates PhD
patents patents lights 2010 2010 2010
ATT 32.77*** 11.82*** 6.141*** 30589.2 8212.4 186.6
(10.02) (4.471) (2.110) (31801.1) (8824.0) (234.3)
Anal. S.E. [9.87] [4.36] [2.13] [30957.72] [8188.28] [230.41]
Raw Diff. 37.11 13.57 23.71 79894.37 22509.07 596.18
N. Treated 70 70 70 70 70 70
N. Controls 1399 1399 1399 1399 1399 1399
Notes: Bootstrapped standard errors in parentheses. * = significant at the 10% level, ** = significant at the 5%
level, *** = significant at the 1% level.
20. Introduction Background Data Methodology Results Conclusion
Municipal-level results: Nearest neighbour matching
(1) (2) (3) (4) (5) (6)
All Frac. Night Population Graduates PhD
patents patents lights 2010 2010 2010
ATT 34.47*** 11.55** 7.193** -3087.8 -143.8 -36.21
(10.14) (4.513) (3.641) (43328.9) (11779.2) (331.1)
Anal. S.E. [9.92] [4.40] [3.74] [46529.55] [12294.66] [349.96]
Raw Diff. 37.11 13.57 23.71 79894.37 22509.07 596.18
N. Treated 73 73 73 73 73 73
N. Controls 60 60 60 60 60 60
Notes: Bootstrapped standard errors in parentheses. * = significant at the 10% level, ** = significant at the 5%
level, *** = significant at the 1% level.
21. Introduction Background Data Methodology Results Conclusion
Firm-level results
Solow
Residual
Labor
Productiv.
R&D
Dummy
Innovation
Dummy
Product
Innovation
Dummy
Process
Innovation
Dummy
Innovation
Share of
Sales
(1)
ATT -0.019 0.068 -0.014 -0.020 -0.021 -0.001 0.280
(0.025) (0.096) (0.024) (0.033) (0.028) (0.027) (1.008)
M 1494 1517 2154 2154 2154 2154 2113
(2)
ATT -0.014 0.035 -0.007 -0.011 -0.021 0.010 0.477
(0.030) (0.115) (0.027) (0.037) (0.032) (0.030) (1.124)
M 1205 1222 1711 1711 1711 1711 1676
(3)
ATT 0.031 0.141 -0.061 -0.039 -0.007 -0.001 -0.854
(0.038) (0.168) (0.049) (0.040) (0.033) (0.039) (1.685)
M 401 404 543 543 543 543 534
(4)
ATT 0.041 -0.282 0.036 0.009 0.027 -0.009 -2.014
(0.051) (0.332) (0.060) (0.049) (0.037) (0.032) (2.278)
M 74 76 99 99 99 99 98
Notes: * = significant at the 10% level, ** = significant at the 5% level, *** = significant at the 1% level. ATT
effects are calculated via weighted OLS, with weights associated to CEM-imputed strata, and by restricted sample
to observations belonging to matched stata following Iacus, King and Porro (2009, 2011). The size of the restricted
sample is denoted as M. Standard errors are clustered by municipality. Each set of results (1)-(4) is obtained by
applying a different CEM algorithm:
(1) CEM on all firm-level characteristics Zic (size, age, foreign and/or state controlled, market scope – local,
national or international, manufacturing/retail dummy, % of university educated employees);
(2) CEM on all firm-level characteristics Zic , on modern municipal characteristics in Xc (fractional patents,
population, PhD population) and on the number of R&D institutions from Soviet times;
(3) CEM as above in (2), forcing exact matching on the larger Russian economic regions;
(4) CEM as above in (2), forcing exact matching on Russian regions (e.g. oblast, krai).
Coarsening measures for the continuous variables in the dataset are available upon request.
22. Introduction Background Data Methodology Results Conclusion
Conclusion: Summary
Science cities are more innovative: they produce more patents,
ceteris paribus
They still present a stronger concentration of R&D activities
despite the withdrawal of much of governmental support
Firms in science cities, however, no more likely to innovate
and do not seem to be more productive or profitable ⇒
understanding why is crucial for policy
23. Introduction Background Data Methodology Results Conclusion
Conclusion: Mechanisms
Mechanism: it is difficult to distinguish between: a. direct
government intervention, b. persistence of human capital, c.
knowledge spillovers
Ideas to motivate a story about a combination of b. and c.:
Low interregional mobility, significant part of worker
compensation paid in kind
Substantial brain drain of scientists, researchers and engineers
after the collapse of the Soviet Union
However, many stayed in Russia, but left science and pursued
other career options, including opening a business
⇒ Knowledge spillovers upon transition to a market economy
Any explanation should address why innovation advantages do
not transfer into productivity or broader economic advantages
(institutional problems: the usual culprit)