The Long Run Effects of R&D Place-based Policies:

Evidence from Russian Science Cities∗

Helena Schweiger†

Alexander Stepanov‡

Paolo Zacchia§

19 May 2017

Abstract

We study the long-run effects of a unique historical place-based policy targeting R&D: the

creation of “Science Cities” in former Soviet Russia. The establishment of Science Cities and

the criteria for selecting their location were largely guided by political and military-strategic

considerations, in addition to economic ones. We compare the current socio-economic

outcomes of Science Cities to those of matched localities that were similar at the time of

establishment. We find that despite massive cuts in government support to R&D following

the dissolution of the USSR, Science Cities still host a more educated population, apply

for more patents, and do better in terms of several proxies of economic development even

though they are not more populous. These findings are consistent with a long-run shift

of the spatial equilibrium due to a combination of barriers to labour mobility and localised

knowledge spillovers. By analysing firm-level data, however, we obtain only mixed evidence

in favour of the existence of spillover effects with a large spatial breadth.

JEL Classification: O33, O38, O14

Keywords: Innovation, Russia, science cities

∗We would like to thank Sergei Guriev, Maria Gorban, Andrei Markevich, and Natalya Volchkova forhelpful discussions as well as the participants at the SITE Academic Conference: 25 years of transition,8th MEIDE conference, 2015 PacDev conference, 17th IEA World Congress, 16th Uddevalla Symposiumand seminars at the EBRD, Higher School of Economics and New Economic School for their commentsand suggestions. Irina Capita, Jan Lukšic and Maria Vasilenko provided excellent research assistance. Theviews expressed in this paper are our own and do not necessarily represent those of the institutions ofaffiliation.

†European Bank for Reconstruction and Development (EBRD). E-mail: schweigh@ebrd.com.‡European Bank for Reconstruction and Development (EBRD). E-mail: stepanoa@ebrd.com.§IMT Lucca Institute for Advanced Studies. E-mail: paolo.zacchia@imtlucca.it.

1 Introduction

The effectiveness of public support to science and R&D is a longstanding issue in the

economics of innovation. Both direct subsidies and indirect incentives to research and

science are usually motivated by the existence of positive externalities (or other types of

market failures) which, in the absence of public intervention, cause underinvestment in

R&D. Some specific innovation policies, like the top-down creation of local R&D clusters,

are characterised by a geographical local dimension. In these contexts it is relevant for

policy to assess the spatial extent of knowledge externalities, which correspond to one

of the three agglomeration forces first identified by Marshall (1890), and that parallel the

learning effect from the more recent classification by Duranton and Puga (2004). In ad-

dition, in these cases the debate about the effectiveness of localised innovation policies

mixes with the one about broader (that is, not innovation-specific) place-based policies.

In particular, it is questioned whether place-based policies have any post-intervention

effect in the long run, in the absence of which their net welfare effect is as likely to be

negative as it is to be positive (Glaeser and Gottlieb, 2008).1

This paper examines the long-run impact of a localised innovation policy: the es-

tablishment of highly specialised Science Cities in the territory of modern Russia during

Soviet times. These cities were middle-sized urban centres created by the Soviet govern-

ment according to a grand strategic plan of technological advancement. They hosted a

high concentration of R&D facilities – often the main economic activity in town – typi-

cally built around a specific technological purpose. Science Cities emerged in the con-

text of technological and military competition of the Cold War; unsurprisingly, most of

them were specialised in military-applicable fields, such as nuclear physics, aerospace,

ballistics and chemistry – although a minority of them were focused in other areas. The

above sectors remain, to this day, those in which Russia maintains a comparative tech-

nological advantage.

1Their argument is based on the interaction between congestion effects and spatial agglomeration ex-ternalities – such as those due to local knowledge spillovers – in a spatial equilibrium model that allows formovement of workers across places. In their theoretical framework, place-based policies that move em-ployment between areas are welfare-improving only if they are effective at shifting economic activity to abetter long-run equilibrium, one in which employment is reallocated in such a way that the increase inself-reinforcing agglomeration forces more than countervails the possibly negative effects from increasedcongestion. Multiple equilibria with such features are however only possible if agglomeration externalitiesfeature non-linearities. This has motivated subsequent empirical research aimed at uncovering agglom-eration effects and their (potential) non-linearities. See also the discussion by Glaeser and Gottlieb (2009)and Kline and Moretti (2014b).

1

While one may question whether the institutional context of Russian Science Cities

is comparable to that of other industrialised countries, this historical experience stands

out with some unique features that motivate its analysis. First, the locations of Science

Cities were typically chosen by the Soviet leadership with criteria that are unusual for a

capitalistic market economy. According to historical research on the topic by Aguirrechu

(2009), since the Soviet government had the power to allocate both physical and human

capital where it deemed necessary, the potential for economic development and local

human capital accumulation was typically not, at the margin, a determinant of a loca-

tion’s choice for the establishment of a Science City. Instead, between any two places

that were geographically close and similar in other respects, the choice usually fell over

the one offering more potential for secrecy and safety from outside, foreign interference

(in the form of R&D espionage). This greatly diminishes concerns for selection biases

due to unobserved determinants of future development, which typically affect studies

about innovative clusters in other countries.

Second, the transition to a market economy that followed the dissolution of the So-

viet Union resulted in a large negative shock for Russian R&D, as direct governmental

expenditure in R&D as a percentage of GDP fell by about 75 per cent. Coupled with the

fact that in those years the Russian private sector was only in its infancy, this resulted

in more than half of the scientists and researchers of post-1991 Russia losing their job.

Unfortunately, adequate information about funding directed to former Science Cities

during the transition is not publicly available, but we argue that as a consequence of

the shock many of them were drastically downsized. In fact, only in the early 2000s was

state support officially resumed for some of the former Science Cities: fourteen in total,

which today bear the name of Naukogrady. Together, these developments indicate that

both the initiation and discontinuation of the Science Cities policy were largely driven

by unique factors, mainly orthogonal to determinants of current demographic and eco-

nomic conditions. In addition, controlling for current governmental support enables

us to evaluate to what extent our results depend on it, instead of being caused by other

factors that relate to persistent effects from the Soviet-era policy.

We estimate the effect of the past establishment of a Science City on the follow-

ing present characteristics of Russian municipalities: innovation (patent output), hu-

man capital (university-graduate or PhD-graduate share of the population) and various

proxies of economic development. In order to give a causal interpretation to our esti-

mates, we construct appropriate control groups by employing matching techniques –

2

propensity-score and Mahalanobis matching. In particular, we match Science Cities to

other localities that, at the time of selection, were similar to them in terms of character-

istics that could affect both their probability of being chosen and their future outcomes.

Our main identifying assumption is that, conditional on these variables, the choice of

a locality was determined at the margin by factors – such as the potential for secrecy

– that would be independent from future, post-transition outcomes. In order to im-

plement this strategy we construct a unique dataset of Russian municipalities (rayons),

which combines both historical and recently observed characteristics.

Furthermore, we complement municipality-level empirical analysis with an addi-

tional set of estimates using firm-level data. We use Russian firm-level data from the

fifth round of the Business Environment and Enterprise Performance Survey (BEEPS V),

sampled from regions with Science Cities locations. Using a variety of specifications,

we evaluate to what extent the distance of a firm from a Science City correlates with

its outcomes in terms of innovation and productivity. BEEPS V is particularly useful in

this regard, as it features an innovation module with detailed information about recent

innovative activities by firms. This analysis is meant to evaluate if, in the modern Rus-

sian economy, the effect of Science Cities continues to spill over on other firms that are

located nearby, and to what economic and geographical extent.

Our results can be summarised as follows. In today’s Russia, Science Cities from the

Soviet era apply for more patents, are more economically developed and host a more

educated population than localities that were comparable to them at the time when the

program started. In line with our expectations (motivated by the specific institutional

context), the estimated treatment effect is lower than the raw sample mean difference

for all variables except patent applications, in whose case no ex-ante bias can be at-

tested. In addition, when we restrict the analysis to localities that have no official sup-

port by the present Russian government, our results remain largely unchanged, with one

exception: for patent outcomes whose point estimates decrease by about 60 per cent.

This indicates that long-run persistence of knowledge capital – possibly coupled with

knowledge spillovers – is likely the most relevant mechanism driving the results. This is

in line with the findings from a related study about the long-run effects of Soviet R&D,

which exploits regional variation of R&D employment at the higher administrative level

of Russian oblasts (Ivanov, 2016). The firm-level analysis, on the other hand, provides no

robust evidence that firms benefit from locating close to Science Cities. While further

research may support stronger conclusions, this is suggestive of the fact that spillover

3

effects due to Science Cities (if they exist at all) are limited in scope.

Our paper contributes to various strands of the literature. First, we add to the set of

studies about the evaluation of place-based policies; for a recent survey of the empirical

research see Neumark and Simpson (2014). The majority of these papers analyse spe-

cific policies enacted in the US (Neumark and Kolko, 2010; Busso et al., 2013; Kline and

Moretti, 2014a) or in the EU (Bronzini and de Blasio, 2006; Criscuolo et al., 2012; Givord

et al., 2013). Among the few that focus, like us, on a non-western country, there is a no-

table contribution by Wang (2013) about Chinese Special Economic Zones (SEZs). The

empirical challenges faced by these studies are typically about constructing appropriate

control groups, and disentangling direct effects from spillovers. Methodologically, our

paper is most directly related to the study by Kline and Moretti (2014a) on the Tennessee

Valley Authority; like in their study, we apply a matching strategy in order to uncover the

long run consequences of our policy of interest. Unlike Kline and Moretti, however, we

find that these are not confined to the sector directly targeted by the policy – that is, R&D

– arguably because of the effect of knowledge spillovers.

Second, and relatedly, we contribute to the more general search of agglomeration

effects – and in particular of the third Marshallian force, localised knowledge spillovers

– in urban and regional economics. This has long been a traditional field of investiga-

tion for economic geographers, with a particular interest in innovation clusters. Follow-

ing seminal contributions by Jaffe (1989), Glaeser et al. (1992), Audretsch and Feldman

(1996) and others, a large literature has developed.2 Recently, the issue has caught the

attention of economists working in more diverse fields. Moretti (2004) shows that in US

cities, the level of education of the workforce affects firm productivity across sectors.

Ellison et al. (2010) simultaneously test all three Marshallian theories by looking at the

co-location of plants across industries. Greenstone et al. (2010) demonstrate the exis-

tence of local productivity spillovers following the opening of a so-called Million Dollar

Plant. In two separate contributions, Bloom et al. (2013) and Lychagin et al. (2016) find

an association between firms’ R&D spending and the productivity of those nearby.3

2We propose two fairly recent surveys: Beaudry and Schiffauerova (2009) focus on the Marshall vs.Jacobs debate around the prevalence of, respectively, within- versus between-industry local knowledgespillovers, while Boschma and Frenken (2011) devote special attention to studies within the evolutionaryeconomic geography research agenda.

3Other, related studies discuss to what extent patent citations can be exploited to recover patterns oflocalised knowledge spillovers. The original contribution having documented this association (Jaffe et al.,1993) has been criticised on methodological grounds in the paper by Thompson and Fox-Kean (2005).Later, Breschi and Lissoni (2009) attempted to bridge the gap between the contending parties of that

4

The specific institutional setting of this paper relates it to other, somewhat diverse

contributions about the consequences of historically massive forms of government in-

tervention on long-run economic and technological development, either in Russia or

elsewhere. Cheremukhin et al. (2017) argue that the “Big Push” industrialisation policy

enacted in the USSR under Stalin did not succeed in shifting Russia onto a faster path

of economic development. In addition, Mikhailova (2012) finds negative welfare effects

from the regional demographic policies enacted by the Soviet Union. However, the pic-

ture looks different in the more specific case of R&D policies. The already cited contribu-

tion by Ivanov (2016) demonstrates how Russian regions with more R&D personnel be-

fore the transition do better today at expanding employment in high-tech sectors. Out-

side Russia, Moretti et al. (2016) show that in OECD countries increases in government-

funded R&D for military purposes have positive net effects on TFP, despite crowding out

private expenditures in R&D.

This paper is organized as follows. Section 2 provides a brief account of the historical

experience of Soviet Science Cities. Section 4 describes the data employed in both the

municipal- and firm-level analysis, along with summary statistics. Section 5 outlines our

empirical methodologies. Sections 6 and 7 discuss the empirical results, respectively

for the municipal-level and the firm-level analyses. Finally, Section 8 summarises the

findings and concludes the paper.

2 Historical and Institutional Background

2.1 History of Science Cities

In the context of the Cold War, geopolitical competition between the USA and the USSR,

the Soviet leadership prioritised allocation of the best resources – including human ones

– to sectors considered vital to the country’s national security. Unsurprisingly, around

two-thirds of all Soviet R&D spending was set for military purposes, and almost all of

the country’s high-technology industry was in sectors directly or indirectly related to de-

fence (Cooper, 2012). Science Cities emerged in this environment. They were 95 middle-

sized urban centers which the Soviet government endowed with a high concentration of

research and development facilities, and – in analogy with many other one-company

debate by showing that the geographical concentration of patent citations is largely due to the spatiallocalisation of co-authorship networks.

5

towns that were typical of the Soviet socialist economy (see Commander et al. (2011)) –

they were devoted to a particular scientific and technical specialisation.4

The former Soviet Union was in a way a pioneer in public investment in science

and in place-based policies that focused on R&D. The model of innovation followed

by the Soviet authorities since the early 1930s was the creation of “special-regime en-

claves intended to promote innovation” (Cooper, 2012).5 Proper Science Cities began to

develop around strategically important (military) research centres from the mid-1930s,

and more of them followed after the Second World War. Table A.1 in Appendix A pro-

vides an overview of their location, the year in which they obtained Science City status

and their priority specialisation areas; this information is based on Aguirrechu (2009),

Lappo and Polyan (2008), NAS (2002) and Russian Wikipedia articles.6 Figure 1 maps

the location of Science Cities: almost all of them were located in populated parts of Rus-

sia, with a high concentration in two specific areas (the Moscow region and the Urals).

As they specialised in industries with high technological intensity, Science Cities

needed access to suitable equipment, machinery, energy supply, intermediaries and

qualified personnel. With the purpose of encouraging skilled individuals to relocate to

Science Cities, the Soviet government strove to make the living conditions in these lo-

calities better than the Soviet standard, by making available to residents a wider choice

of retail goods, more comfortable apartments as well as more abundant cultural oppor-

tunities than elsewhere in the country. Typically, the urban characteristics of planned

Science Cities did not match those of contemporary settlements. In fact, while about

two thirds of Science Cities were established by repurposing existing towns, the rest were

built from scratch, with the specific aim of co-locating scientific research centers, train-

ing institutes and manufacturing facilities (Aguirrechu, 2009).

4The term Science City, Naukograd in Russian, was first introduced in 1991 (Ruchnov and Zaitseva,2011). The former Soviet Union was not a Science Cities pioneer — the first Science City was establishedin 1937 in Peenemünde, Germany (Ruchnov and Zaitseva, 2011) — but it implemented the idea to a muchlarger extent.

5These enclaves first appeared as secret research and development laboratories (so-called Experimen-tal Design Bureaus or more commonly, sharashki) in the Soviet Gulag labour camp system. Scientistsand engineers employed in a sharashka were prisoners picked from various camps and prisons, as-signed to work on scientific and technological problems — a task at which they were quite successful.Famous sharashka inmates included Sergey Korolyov (aircraft and rocket designer); the rocket engine de-signer Valentin Glushko; aircraft designers Andrei Tupolev, Vladimir Petlyakov, Vladimir Myasishchev andLeonid Kerber; the geneticist and radiobiologist Nikolay Timofeev-Ressovsky; the inventor of the straight-flow boiler Leonid Ramzin; and the pioneer of aeronautics and space flight Yuri Kondratyuk.

6A definitive list of Science Cities does not exist, but we are confident that the sources we used are thebest available.

6

Starting in the 1940s, with the need to protect the secrecy of the nuclear weapons

program in the Cold War environment (Cooper, 2012), Science Cities whose main objec-

tive was to develop nuclear weapons, missile technology, aircraft and electronics were

closed in order to maintain security and secrecy.7 This had several implications: non-

residents needed explicit permission to travel to closed cities and were subject to doc-

ument checks at security checkpoints; foreigners were prohibited from entering at all;

relocating to a closed city required a security clearance by the KGB; and dwellers had to

keep their place of residence secret.8 Closed cities that specialised in R&D were often

located in remote areas situated deep in the Urals and Siberia – out of reach of enemy

bombers – and were represented only on classified maps.

Another subset of Science Cities stands out: the so-called academic towns (akadem-

gorodki), which are comparable to U.S. university campuses of large dimension. In fact,

academic towns were semi-isolated neighborhoods of a larger city, endowed with R&D

facilities, housing for R&D staff and their families, as well as basic local infrastructure

such as shops, schools and small hospitals. In all other respects academic towns were

dependent on the larger city they were attached to; in this sense, unlike other Science

Cities, they were not self-sufficient entities. Academic towns appeared after the Second

World War to the east of the Urals, as part of the Soviet program to develop Siberia. The

first and largest academic town was founded in the outskirts of Novosibirsk in 1957.9

2.2 Location of Science Cities

Thanks to the cited historical research by Aguirrechu (2009), however, it is possible to

identify some general factors that drove the choice of locations for specific groups of Sci-

ence Cities. Two general themes emerge. First, the relevant natural, socio-economic and

demographic factors that influenced the choice of a place usually varied by the specific

function of a Science City. Second, at the margin the choice of a location over another

7Examples of closed Science Cities include several ROSATOM (State Atomic Energy Corporation) cities,such as Ozyorsk, Snezhninsk, Zelenogorsk, Protvino and Novouralsk (Table A.1 in Appendix A). All butProtvino are still closed.

8Cities specialising in R&D were only a minority of all Soviet municipalities subject to these restrictions.The majority of closed cities were closed because they hosted strategic military troops or were located nearthe border. We do not consider those in our analysis as they were not Science Cities with a focus on R&D.We include (data permitting) all those Science Cities that were closed because their main objective was todevelop nuclear weapons, missile technology, aircraft and electronics (high-tech sectors).

9Other academic towns were created in Apatity (1954), Irkutsk (1988), Krasnoyarsk (1965) and Tomsk(1972) (Aguirrechu, 2009).

7

usually depended on political, military and security motivations that are arguably unre-

lated with the determinants of economic outcomes in a typical market economy. These

two considerations, on which we expand below, inform this paper’s empirical strategy.

Specifically, our matching strategy rests on the assumption that controlling for certain

relevant factors, Science City status is unrelated to current outcomes.

In terms of socio-economic and demographic characteristics that affected the loca-

tion of Science Cities, the most relevant one that is identified by Aguirrechu is, unsur-

prisingly, the pre-existing level of economic and social development. With some excep-

tions, Science Cities were established in the areas of Russia that were the most industri-

alised, urbanised, and with better educated population; this way, Science Cities could

have easier access to qualified personnel or be able to attract it with minor additional

costs. For this reason, Science Cities are also for the most part located in the western,

warmer part of Russia, within the humid continental climatic region typified by large

seasonal temperature differences. Historically, the socio-economic development differ-

entials between Russian regions strongly correlate with temperature gradients along a

longitudinal axis.10

Other geographical factors differ by Science City type. Those engaging primarily in

basic R&D – such as Dubna and Troitsk – were typically semi-isolated, to be found ei-

ther in outer parts of a region or in the territories between major highways and rail-

roads. Science Cities engaging primarily in applied, production-oriented R&D in civil-

or double-purpose industries (such as electronics or aviation), by contrast, were located

either close to the regional capital or in the proximity to transportation links: with a very

Marshallian motivation, these cities were in more need of easy access to both upstream

suppliers and downstream “buyers” (a term to be interpreted in the context of a social-

ist economy). Heavy industry and nuclear technology needed large amounts of water,

therefore Science Cities specialised in those areas were typically built close to rivers or

lakes. For analogous reasons, Science Cities focused in military shipbuilding – such as

Severodvinsk – had to be located on the coast.

The exact location of a Science City, however, often depended on very idiosyncratic

factors whose main motivation was military, political or strategic. In general, Aguirrechu

underlines the fact that, whenever a Science City had to be set in an urbanised and rela-

tively developed region, the choice between any two similar localities usually fell on the

10In Russia, temperature changes more along the west-east axis, than along the north-south axis; thus,for two localities with the same latitude, the eastern one is typically colder.

8

one with the most potential to maintain secrecy and minimise the threat of spying; he

supports this argument with anecdotal evidence. In this respect, it is not surprising that

many Science Cities were established in the proximity of Moscow, close to the central

government and the headquarters of security agencies such as the KGB. At the extreme,

considerations of this kind overrode all the others. In particular, Science Cities special-

ising in some applied R&D fields such as the production of nuclear and strategic arms

faced a much higher threat of bombing and spying, and were located in regions far from

the borders and in poorly populated municipalities far from the regional centre (with

limited transport links). Examples include Sarov and Snezhinsk.11

Some of these idiosyncratic factors depended on other historical and political cir-

cumstances. Following the evacuation of factories from the European part of the Soviet

Union beyond the Urals during the Second World War, those areas developed rapidly.

On the one hand, this may explain the concentration of many Science Cities in the Urals

area. On the other hand, this was the main motivation for the establishment of academic

towns in Siberian regions with rising industrial and strategic importance but limited sci-

entific capacities, in order to foster their industrial development. There is evidence that

the research in natural sciences conducted in these academic towns was directly linked

to the specific issues faced by Siberia (Aguirrechu, 2009).

2.3 R&D and Science Cities after the Collapse of the Soviet Union

Following the dissolution of the USSR, Russia underwent a profound and for the most

part painful transformation from a planned to a market economy. The withdrawal of the

state from many industries dramatically affected the R&D sector. As shown in Figure 2,

the end of the Soviet Union in fact resulted in a significant drop in R&D funding. From

the 1990 level of about 2 per cent of GDP, which is close to the OECD average, Russian

11These two places provide a particularly indicative example of idiosyncratic factors affecting the loca-tion of Science Cities: sometimes, this was determined by the presence of other Science Cities, or lackthereof. Specifically, Snezhinsk (Chelyabinsk region) was established as a double of Sarov (Nizhny Nov-gorod region) with the main purpose of keeping the industry working even if one of the two places weredestroyed, but also to create inter-City competition. As Sarov is located in a relatively remote location inthe European part of Russia, Snezhinsk had to be placed in a similarly out-of-reach area, but to the Eastof Urals. Officials reportedly considered locations in Novosibirsk, Tomsk and Krasnoyarsk regions, but ul-timately decided on Chelyabinsk and Snezhinsk in particular because of the proximity of another ScienceCity, Ozyorsk, which could supply inputs to Snezhinsk. A similar pattern of interplay between decisionsaffecting different Science Cities was not unique; for example, the four places specialized in the produc-tion of enriched uranium were also located far from each other: respectively in the Sverdlovsk, Tomsk,Krasnoyarsk and Irkutsk regions.

9

gross R&D expenditures (GERD) fell to a mere 0.74 per cent in just two years.12 This is

even more dramatic in face of the fact that the Russian GDP shrank significantly in the

initial years of transition, for a total decline by about 50 per cent between 1990 and 1995

according to official government statistics.

The steep fall in GERD was associated with a gradual decrease in the number of

R&D personnel of similar relative cumulative magnitude, as shown in Figure 3. Whereas

in Soviet times the wages of scientists were 10-20 per cent higher than average, they

dropped to 65 per cent of the average wage as early as 1992, following the withdrawal

of the state from the R&D sector (Saltykov, 1997). Even worse, during the 1990s many

scientists did not even receive their salary or received only a fraction of it over extended

periods (Ganguli, 2014). Unsurprisingly, this stimulated the outflow of scientists to the

business sector: 71 per cent of people who left the R&D sector in those years list higher

salaries as the main reason for changing employment. Low remuneration was not the

only reason for researchers to leave the R&D sector: with the removal of previous restric-

tions to individual mobility, scientists were allowed to migrate. However, even according

to the less conservative estimates, emigration accounted for less than 3 per cent of the

total outflow of staff from the R&D sector,13 and was more driven by ethnic factors than

by differential labour market conditions (Gokhberg, 1997).

It is difficult to identify the exact extent to which this shock affected Science Cities.

Lack of consistent municipal or enterprise-level data, aggravated by the secrecy of many

Science Cities at the beginning of transition, impedes a comprehensive analysis. We ar-

gue, however, that dissolution of the USSR implied a massive downsizing, if not suspen-

sion, of the Soviet Science Cities program. Two facts point to this conclusion. First, most

Science Cities were concentrated in military-related R&D, which in aggregate terms un-

derwent a similar decline as the overall GERD in post-1991 Russia. Specifically, in 1991

12The collapse, which was mainly caused by a decline in government spending, led to a dramatic changein the distribution of Russian GERD by source. Whereas in 1991 the government directly funded 95 percent of total R&D in Russia, by 1995 its share shrank to 60.5 per cent. Other sources became more im-portant, namely the enterprise sector, foreign funding and own funds of R&D institutions. Overall, thisstimulated competition within the sector and forced R&D institutions to expand to such R&D-related ac-tivities as consulting and information services.

13Graham and Dezhina (2008) cite estimates of a number of researchers between 7,000 to 40,000 leavingRussia in the period from 1993 to 1996. Using official statistics, Gokhberg and Nekipelova (2002) estimatethat approximately 20,000 individuals working in the “Science and Scientific Services” sector emigratedduring the 1990s, although many of these individuals may not have been research scientists. Gokhberg(1997) cites a study by CSRS finding that emigrants accounted for only 0.5 per cent of the total outflow ofstaff from the R&D sector. He also estimates the emigrations rates of Russian scientists between 1992 and1999 to be between 0 and 2.4 per cent.

10

defense R&D accounted for 0.75 per cent of Russian GDP, eventually falling to about

0.2 per cent – a figure comparable with the OECD average but half that in the US. Af-

terwards, it has been fluctuating around this level until the early 2000s, when a steady

increase can be identified (see Figure 4). Second, the Russian government has, starting

in the early 2000s, re-established direct support for selected previous Science Cities. In

total, only 14 localities bear the official status of Naukogrady in today’s Russia. This re-

flects the choice to resume the previous system of public R&D support, albeit to a much

smaller scale and in a more concentrated fashion.

3 Theoretical Hypotheses

In light of the historical and institutional background described above, we present a

few hypotheses about the long-run consequences of the Soviet-era Science Cities pro-

gram that we then test statistically. To motivate them, we implicitly refer to a theoretical

framework à la Glaeser and Gottlieb (2008); in that context, non-linearities of urban ag-

glomeration externalities (here, localised knowledge spillovers) would induce a change

in the spatial equilibrium and thus persistent effects even after the discontinuation of

the policy (the past R&D investment).

Municipal-level hypotheses

HM.1 Municipalities with Science Cities have retained a higher stock of knowledge cap-

ital than similar municipalities without Science Cities. While the former enjoyed a

higher concentration of knowledge capital than the latter during the Soviet times,

persistent effects in the modern Russian market economy may be due to either:

(i) a change in the spatial equilibrium due to agglomeration effects; (ii) frictions to

internal mobility and emigration after the collapse of the Soviet Union (resulting

in many researchers staying put), (iii) continued governmental support for local

R&D. The aggregate stock of knowledge/human capital is proxied by the share of

population with a university degree or a PhD.

HM.2 Municipalities with Science Cities are more innovative than similar municipali-

ties without Science Cities. This would be a consequence of HM.1, as cities with

higher knowledge capital are more likely to engage in innovative activities. We

proxy the latter by the number of patents applied for at the foreign patent offices.

11

HM.3 Municipalities with Science Cities are more economically developed than simi-

lar municipalities without Science Cities. This, in turn, would follow from HM.1,

because of aggregate returns to knowledge capital coupled with spillover effects,

as well as from HM.2, because of returns to innovations. We use several proxies

for the level of development: nighttime lights as well as the number, density and

labour productivity of SMEs (in total economy and in manufacturing only).

Firm-level hypothesis

HF.1 Firms located closer to Science Cities are more innovative and better performing

than firms located further away. Because of HM.1, Science Cities may spur innova-

tion and productivity effects in firms located nearby due to knowledge spillovers.

We look at whether firms closer to Science Cities are more likely to engage in any

R&D activity, introduce a new product or process, whether they have a larger share

of their sales due to their innovations, or are more productive.

4 Data and Descriptive Statistics

We evaluate the long-run effects of Science Cities by employing a unique dataset, which

contains information previously unavailable in electronic format. Specifically, it com-

bines: (i) our database on Science Cities, which is described in Section 2 and reported

in Appendix A; (ii) municipal-level data that aggregate various sources about historical

and current characteristics of Russian cities; and (iii) a firm-level database that is ob-

tained by merging BEEPS V Russia and Bureau van Dijk’s Orbis data. In what follows, we

describe the latter two.

4.1 Municipal-level Data

We construct a municipal-level data set for all Russian municipalities at the rayon ad-

ministrative level.14 We obtain administrative data from official sources, and we merge

municipalities to different types of data with the support of GIS software. We manually

assign Science City status to each municipality; as some of these include more than one

14In this paper, we use the English term “municipality” to denote the corresponding Russian rayon ad-ministrative level, and the word “region” to refer to any area at the higher oblast, krai or republic level. InRussia, these three administrative levels do not overlap.

12

historical Science City, however, in total the data include 88 municipalities with at least

one Science City.15 In a few cases historical and current municipal boundaries do not

match exactly, thus we check our data manually. Data types and sources are described

in more detail in Appendix B; here we briefly summarise them distinguishing between

current socio-economic outcomes, historical and geographical characteristics.

Current Outcomes. Our variables of interest on current characteristics of Russian

municipalities match our hypotheses HM1-3 outlined in Section 3. Specifically, we ex-

tract data about the overall municipal population, the share of this population that at-

tained higher education qualifications, and the share of the population holding a doc-

toral degree (Ph.D. equivalent) from the 2010 Russian Census. We proxy innovation by

the number of patents with a local author address that were applied for at the European

Patent Office (EPO) between 2006 and 2015. Finally, as accurate GDP data at the rayon

level is unavailable in Russia, we use several proxies for economic activity: average light

emitted into space at night (night lights) in 2009-2011,16 as well as a number of vari-

ables about small and medium-sized enterprises (SMEs) from the 2010 SME census by

the Russian Statistical Office (ROSSTAT).17 Regarding the latter, we look at the overall

number, the density and the labour productivity of SMEs.

Geographical Characteristics. We collect or calculate municipality-specific infor-

mation about several geographical characteristics: municipal area, average altitude, as

well as average temperatures in January and July. As Science Cities specialised in some

heavy industries needed to be located close to large amounts of water, we also collect

data on each municipality’s access to the coast, to a major river or to a major lake. We

code these both as dummy variables and as distances between the municipality’s geo-

graphical centroid and the closest source of water in question.

Historical Characteristics. Most Science Cities needed access to good transporta-

15NAS (2002) lists four Science Cities for which only their Soviet-era nomenclature is publicly available:Krasnodar-59, Novosibirsk-49, Omsk-5 and Perm-6. Their exact location is still unclear; thus we excludethese four places from the analysis as they cannot be matched to any municipality. In addition, three pairsof Science Cities are located within the same municipalities (Lyubertsi and Tomilino in Lyuberetsky rayon,Avtopoligon and Orevo in Dmitrovsky rayon, and Peresvet and Remmash in Sergiyevo-Posadsky rayon).Hence, 91 Science Cities are mapped to 88 municipalities with at least one Science City.

16Night lights can plausibly be used as a proxy for economic activity under the assumption that lightingis a normal good; see Donaldson and Storeygard (2016). Examples of papers using night lights as a proxyfor economic activity within geographic units for which no alternative data source is available includeHodler and Raschky (2014) and Storeygard (2016).

17According to the official documents, the study was a full-scale federal statistical observation. However,since all operating SMEs were required to participate we consider it as a census.

13

tion links, while others had to be located in remote areas far from espionage threats. To

account for both factors we use GIS data on Russian railroads in 194318 as well as on the

USSR borders; we code them both as dummy variables and as distances from the mu-

nicipal geographic centroid. To control for the existing level of industrial development

in a municipality, we use information about the number of plants (factories, research

and design establishments) of the Soviet defence industry in 1947 from Dexter and Ro-

dionov (2016). Furthermore, to make sure that we match municipalities of similar size

in terms of population, we use data from the first post-World War II census held in the

Soviet Union, conducted in January 1959. We would prefer to use population data from

the 1940s but there was no census conducted until 1959; moreover, World War II affected

the Russian demography so much that any figures collected before 1941 are inadequate.

Since the 1959 census does not have data on population with higher education at the

municipal level, we use data on the number of higher education institutions and R&D

institutes located in a municipality in 1959 as a proxy for it.

Summary Statistics. Table 1 displays summary statistics for the municipal-level

characteristics and outcomes, distinguishing between municipalities with Science Cities

and all other Russian municipalities. It shows that, on average, Science Cities were lo-

cated in more populous and warmer places, with a higher historical concentration of in-

dustrial plants, universities, and R&D institutes. In addition, all our “current” outcome

variables register positive and significant differences. The empirical analysis is meant to

evaluate if these still hold upon comparing Science Cities to localities that were similar

to them at time of their establishment.

4.2 Firm-level Data Sources

To perform our firm-level analysis, we use the fifth round of BEEPS merged with Bureau

van Dijk’s Orbis database – both for Russia only. BEEPS is a firm-level survey conducted

by the European Bank for Reconstruction and Development and the World Bank. It is

based on face-to-face interviews with managers of registered firms with at least five em-

ployees. Stratified random sampling is used to select eligible firms to participate in the

18In the Soviet economy, railroads were the workhorse of the transportation network; road transportplayed only a secondary role (Ambler et al., 1985). Most of the railroads’ construction took place in tsaristRussia; even in Soviet times railroads were not important just for transportation and mobility, but also asdrivers of regional industrialisation. Using information about the railroad network in 1943 is preferable tolater dates, because the Soviet rail transport became one of the most developed in the world after WorldWar II, driven by the country’s need to extract – and transport – its natural resources.

14

survey. Strata are defined by sector (typically manufacturing, retail and other services),

size (5-19, 20-99 and 100+ employees) and regions within a country. The fifth round of

the survey (BEEPS V) was completed in Russia in 2012 with 4,220 top managers (CEOs)

across 37 regions, with at least one in each federal district. The database contains GPS

coordinates of the firm’s location, based on which we can determine distances from Sci-

ence Cities. Additional information about BEEPS V Russia can be found in Appendix C.

Outcomes. BEEPS V included, for the first time, an Inovation Module. This lets us

look at whether in the last three years prior to the survey the firms engaged in R&D,

introduced a new product, process or technological innovation and the share of their

sales due to newly introduced products. We manually clean the information contained

in the innovation module – that is, for each firm we verify whether survey responses

match the firm’s main product and industry, and compare the descriptions of the main

new product or process reported in the survey with the definitions given in the Oslo

Manual (OECD and Statistical Office of the European Communities, 2005), reclassify-

ing those that do not match.19. In addition, we are able to match about 75 per cent of

BEEPS V firms to Orbis accounting data, and thus obtain officially reported measures

of economic performance for a subset of firms (rather than self-reported ones in BEEPS

V). Specifically, we use the log of raw labour productivity as well as the log of operating

revenue; the latter is used as a proxy of total output in a reduced form model akin to a

Cobb-Douglas production function.

Controls. BEEPS V Russia also contains measures for several firm characteristics,

such as: age, industry, exporter status, ownership, geographical scope of the main mar-

ket, exporter status (direct or indirect), the number of permanent, full-time employees,

as well as the share of employees with a university degree. We include them on the right-

hand side of all our empirical models with an innovation-related outcome. For regres-

sions with log labour productivity we omit the employees variable; for regressions of log

operating revenue we include the logs of all “conventional inputs” (number of employ-

ees, total assets and the cost of materials).

Summary Statistics. Table 2 reports descriptive statistics at the firm level. Notice-

ably, more than a fifth of firms (21.6 per cent) reports at least some type innovation in

the last three years prior to the survey; although the fraction of those performing R&D

is much lower (11.1 per cent). Innovative products on average generate 3.9 per cent of

19More details about the cleaning process can be found in Appendix C

15

total sales.

5 Empirical Methodology

5.1 Municipal-level Analysis

In order to test our hypotheses HM.1-3 we compare the long-run outcomes Yi q of mu-

nicipalities with Science Cities with those of municipalities without Science Cities (so-

called “ordinary” municipalities) that are similar to the former in terms of geographical

and historical characteristics Xi k . Here, i = 1, . . . , N stands for municipalities, q = 1, . . . ,Q

indexes different long-run outcomes, and k = 1, . . . ,K different physical and historical

characteristics (covariates).

Our identifying assumption in estimating the long-run impact of Science Cities - the

average treatment effect on the treated (ATT) - is that, conditional on the observed ge-

ographical and historical characteristics, the establishment of Science Cities does not

depend on factors that would affect future outcomes:

Yi q (1) ,Yi q (0)⊥Di∣∣ Xi 1, . . . , Xi K (1)

where Di is equal to one for municipalities with Science Cities and zero otherwise, while

Yi q (1) and Yi q (0) correspond to the long-run outcome Yi q for, respectively, Di = 1 and

Di = 0.

In this setting, the rationale of the Conditional Independence Assumption (1) is pro-

vided by the discussion about the factors affecting selection of Science Cities locations

in section 2.2. In particular, between any two otherwise similar locations, the choice of

the Soviet leadership seemed to be guided more by strategic and military considerations

rather than economic ones, as one would otherwise reasonably expect in the context of

a market economy. Nevertheless, we control for a number of factors that might have af-

fected both the choice of the Soviet government and future outcomes, such as: a munic-

ipality’s population; the number of industrial plants, R&D institutes and universities at

the onset of the Cold War; access to transportation links (distance from railroads); access

to major water sources (distance from the coast, major lakes or major rivers); likelihood

of foreigners’ intrusion (distance from the USSR borders) and climate characteristics

(temperature in January and in July, average altitude in the municipality’s territory).

16

We estimate the ATT via nonparametric matching estimators obtained following two

different procedures. Our main results are obtained via the propensity score matching

(PSM) technique: a method for constructing a control group by associating – in this case

– to each municipality with a Science City a number of ordinary municipalities that are

similar to them in terms of the propensity score p (xi ). The latter is, in turn, an esti-

mate of the probability of hosting a Science City for a municipality with characteristics

xi = (Xi 1, . . . , Xi K ); the usual motivation for PSM is that for any two groups with the same

distribution of p (xi ), the joint distribution of all the covariates must also be the same.

We estimate p (xi ) by employing binary outcome models; we assign to each treated unit

one nearest neighbour (with replacement) and then we ensure that covariates are bal-

anced – that is, that their empirical distribution is similar – between the treated and the

constructed-control groups.

A concern with the PSM technique is that it may associate municipalities with Sci-

ence Cities to ordinary municipalities from more distant regions of Russia, thereby mak-

ing the two groups less comparable. We tackle this issue in two ways. First, we include a

flexible polynomial of municipal coordinates (possibly including interactions between

latitude and longitude) in the specification of the propensity score model. As Science

Cities tend to concentrate in specific areas of Russia, this approach should help con-

struct a control group that is located in the same region or nearby. An alternative is to

employ the Mahalanobis matching technique, by which a treated unit i and a control

observation j are matched if their Mahalanobis Distance:

mi j(xi ,x j

)= (xi −x j

)TΣ

(xi −x j

)(2)

(where Σ is the empirical covariance matrix of x) is small enough. Specifically, we as-

sign one nearest-neighbour control (with replacement) to each municipality with a Sci-

ence City, where treated-control distance is given by mi j . We include municipal coordi-

nates into the vector xi , to ensure that municipalities with Science Cities are matched to

“close” places.

Relative to PSM, however, Mahalanobis matching has its own specific drawbacks: in

particular, it is known to perform worse with a high number of covariates, or when co-

variates are not normally distributed (Gu and Rosenbaum, 1993; Zhao, 2004). In order to

address this, we match on the logs of covariates with highly asymmetric empirical dis-

tributions (average altitude, historical population, number of plants, universities and

17

R&D institutes).20 Moreover, instead of matching on measures of the distance of the

municipality from railroads, rivers and lakes, we force exact matching on corresponding

dummy variables denoting whether any railroad, river or lake was actually present in a

given municipality. We expect that, if all these measures are effective at removing con-

cerns for violations of the Conditional Independence Assumption (1), both propensity

score and Mahalanobis matching should return similar estimates of the ATT for all our

outcomes of interest.

The sample coverage differs across outcome variables. Specifically, SME data are

missing for a few regions and municipalities, including some with Science Cities. For

this reason, we implement different propensity score models for non-SME and SME-

based outcomes. In addition, for both PSM and Mahalanobis matching we evaluate

the quality of post-matching covariate balance separately across the two groups of out-

comes. Still, once we associate the q-th outcome of a matched control Ysq (0) to each

treated unit with Ysq (1), we estimate the ATT for all q = 1, . . . ,Q uniformly as:

�AT T q = 1

S

S∑s=1

(Ysq (1)− Ysq (0)

)(3)

where s = 1, . . . ,S indexes successfully matched pairs. Notice that, as our data cover the

universe of Russian municipalities (with the exception of a minority of places for which

SME data is missing), there is no need to apply sampling weights. We calculate stan-

dard errors following Abadie and Imbens (2006), taking into account that the propensity

score is estimated in the case of PSM. In the case of Mahalanobis matching, we calcu-

late estimates of the ATT by applying a “bias-adjustment” formula (Abadie and Imbens,

2011), which is required when matching is performed on more than one covariate.

5.2 Firm-level Analysis

In order to evaluate whether the effect of Science Cities extends beyond the borders

of a specific municipality, and the breadth of their reach (HF.1) we employ firm-level

data. More specifically, we estimate a number of probit models with the following latent

20Clearly, many municipalities featured no plants, universities or R&D institutes at the time when Sci-ence Cities started to be established. In the computation of Mahalanobis distances, we substitute each ofthese covariates Xi k with the corresponding quantity xi k = log(Xi k +1).

18

variable representation:

I∗f r =α+βS∑

s=1exp

[−Dist(

f , s)]+ L∑

`=1γ`W f r,`+ηr +ε f r (4)

where f = 1, . . . ,F indexes firms; s = 1, . . . ,S denotes Science Cities; r is a subscript for

Russian regions (oblast level); I∗f r is the latent variable associated with one specific in-

novation binary outcome I f r ; Dist(

f , s)

is the geodesic distance between the firm f and

Science City s;(W f r,1, . . . ,W f r,L

)is a set of L additional controls; ηr is a region fixed effect;

and finally ε f r is an error term which is distributed as a standard normal.

The main parameter of interest is β, which measures the relationship between firm

innovation and their geographic proximity to Science Cities. Note that exp[−Dist

(f , s

)]expresses exponential decay of a Science City’s influence in space: it is equal to 1 if a firm

is located right in the centre of a Science City, and it is negligible unless firm f and city s

are relatively close to each other. Thus, if a firm is located in a relatively isolated Science

City, the quantity β · φd – where φ f is the standard normal density function evaluated

at the parameter estimates and at firm f ’s values of the RHS variables – approximates

the marginal effect of a Science City on the probability of a positive realisation of I f r for

firm f . This specification is flexible, and it allows for the effect of Science Cities to be

“cumulative” in areas with a high concentration of Science Cities.

In addition, we estimate via OLS a linear version of model (4) for a few continuous

outcome variables:

P f r = α+ βS∑

s=1exp

[−Dist(

f , s)]+ L∑

`=1γ`W f r,`+ ηr +υ f r (5)

where P f r is some firm-specific performance measure. In this case, β is more easily

interpreted as the average change in P f r for firms that are located in a “relatively iso-

lated” Science City. We allow the set of specified controls(W f r,1, . . . ,W f r,L

)to vary across

models depending on the left-hand side outcome of interest.

While we do not attempt to give any causal interpretation to our firm-level results,

we notice that concerns about endogeneity are limited in this setting. Since the creation

of Science Cities predates the establishment of most modern Russian firms – virtually

all in our sample – the only way for the distance-based regressor and the error term to

be correlated is if a Science City attracts or otherwise encourages the location of more

innovative or better performing firms in their proximities. Still, we make no attempts to

19

correct for this possible instance of endogeneity. Our interest, in fact, is about evaluating

in a descriptive sense whether any relationship between Science Cities and firm-level

outcomes extends in space, and we do not intend to remove a potential mechanism by

which such relationships may manifest themselves.

6 Empirical Results at the Municipal Level

6.1 Propensity Score Matching: All Science Cities

We implement two separate propensity score matching (PSM) algorithms in order to ob-

tain corresponding estimates of the ATT for two different sets of outcomes Yi q . The first

set pertains to all the outcomes that are not based on our SME dataset, specifically “sim-

ple” and “fractional” patent applications, census data (population, graduate share, PhD

share) and the lights emitted into space at night as a proxy of economic development. In

this case, PSM is performed on the entire dataset of Russian municipalities. The second

set relates to the outcomes derived from our SME dataset (number, density and labour

productivity of SMEs; both for total economy and manufacturing only). SME informa-

tion is missing for some regions and cities, including some Science Cities; in order to

ensure that the covariates balance on the restricted SME sample, we perform a separate

matching procedure based on a different p-score estimation model.21

6.1.1 Propensity Score Estimation

Table 3 shows the results of the two p-score estimation models respectively for the non-

SME (model 1) and the SME (model 2) groups of outcomes. Both are probit models, but

their specifications differ: either model includes some squared terms for some covari-

ates Xi k that are not included in the other. Model selection is determined by the quality

of covariate balance resulting from matching, on the predicted p-score, via a nearest-

neighbour algorithm with replacement. Both models are estimated on the whole sam-

ple, but in the case of SME outcomes matching is performed on municipalities – both

“treated” and “control” – for which SME information is available.

It is worth discussing some characteristics of the estimated selected models. Both

have strong predictive power, as indicated by the high pseudo-R2 and highly significant

21We notice that the Science Cities for which SME information is missing are former and current closedcities (see section 2). This may explain why these cities are not included in the 2010 SME census data.

20

model χ2 statistics. The probit coefficients associated with most variables are statisti-

cally significant, with some exceptions. The coefficients of both the linear and squared

term of the “distance from railroad” variable are not statistically significant in either

model; a joint χ2 test of significance, however, returns a p-value of 2.4 per cent in model

(1) and a p-value of 1.5 per cent in model (2). In both models, distances from lakes or

from the USSR border are not statistically significant in joint tests. In model (2), the

temperature-related variables are jointly significant with a p-value of 1.9 per cent, but

the number of universities and the number of industrial plants are not statistically sig-

nificant.

6.1.2 Quality of Matching

We match Science Cities to other municipalities via 1:1 nearest neighbour matching on

the propensity score (with replacement) separately for the two sets of outcomes. Be-

cause they fall outside common support, we are unable to match four Science Cities in

the case of non-SME outcomes and two Science Cities in the case of SME outcomes.

Appendix E provides additional details, in the form of figures, about the common sup-

port for the propensity score as well as kernel density estimates of its empirical distribu-

tion (for both treated and control observations, before and after matching, and for both

groups of outcomes). They show that the balancing property is adequately satisfied in

all cases.

Table 4 displays statistics about covariate balance between treated and control cities

in both matched samples. In all cases, standardised biases are below 10 per cent, t-tests

reject the null hypothesis that any covariate is different between treated and controls,

and also the variance ratios denote that matching achieves a remarkably good balance

in the second moments of the covariates’ distributions. Overall, the treated and con-

trol samples differ by about 15 standard deviations in the case of matching for non-SME

outcomes, and by about 23 standard deviations in the case of matching for SME out-

comes.22

6.1.3 ATT Estimation

The PSM-based estimates of the ATT for our twelve outcomes of interest are reported

in Table 5. We summarise the results briefly. Matching confirms that today’s Soviet-era

22These numbers are based on calculations of the Rubin-Thomas B measure (Rubin and Thomas, 1996).

21

Science Cities are more innovative than historically similar localities: between 2006 and

2015 the former have applied for about 28 more patents at the EPO than the latter (10

patents with fractional counting). For both patent-based measures, the ATT does not

statistically differ from the “naive” raw mean difference calculated on the whole sample.

This is not the case, however, for other outcomes of interest. In fact, the 2010 popula-

tion of Science Cities is not statistically different from the one of matched control locali-

ties. On the other hand, Science Cities continue hosting a statistically significantly more

educated population: the college-educated share of their population is 7.4 percentage

points higher, and they also host more PhDs (the share is higher by about 0.3 percentage

points). The college educated share of population estimate is significantly lower than

the naive difference.

In terms of proxies for economic development, our results look as follows. The ATT

estimate for the night lights proxy is positive and statistically significant, with the point

estimate about a third of the size of the raw difference. Upon matching, the cities for

which information about SMEs is available do not seem to have more of these, neither

in absolute nor in relative terms. In fact, the estimated ATT for the absolute number of

SMEs is negative for both total economy and manufacturing, but only the latter is sta-

tistically significant (albeit at 10 per cent significance level). By contrast, raw differences

are positive and statistically significant. In addition, the estimated ATT for the number of

all SMEs relative to population is negative and statistically significant (about four SMEs

less per 1000 people); its equivalent for manufacturing is not statistically significant. Fi-

nally, and interestingly, the estimated ATT of SME labour productivity is positive and

statistically significant both overall and for manufacturing only, with point estimates

about a third lower than naive differences. The results seem to point to an economic

effect of Science Cities operating on the intensive (productivity) margin.

We also perform a sensitivity analysis of our ATT estimates, following Rosenbaum

(2002). Specifically, we simulate the presence of some unobserved factors that would

affect both the outcomes and the probability of receiving the treatment, and we assess

to what extent this would influence our conclusions about the presence of statistically

significant differences in Yi q between treated and (matched) control observations, for

all outcomes q = 1, . . . ,Q. The size of the simulated unobserved factor is given by pa-

rameter Γ≥ 1, which measures the hypothesised odds of receiving the treatment (Γ= 1

in an experimental setting). In Table 5 we report, for each outcome variable, the lower

value Γ∗ that leads to inconclusive tests about the presence of a statistically significant

22

difference between treated and control observations.23 The values of Γ∗ are very high

(above 2) for the patent outcomes, graduate and PhD shares, and for SME labour pro-

ductivity (general); they are higher than 1.5 but lower than 2 for the night lights and SME

labour productivity (manufacturing) outcomes; and below 1.5 for the other outcomes.24

These results are in line with our parametric tests on the estimated ATT parameters, and

demonstrate that our results are very robust to possible threats to identification.25

Finally, it must be mentioned that while our results are based on one-to-one match-

ing, the main qualitative conclusions are not altered in the case of one-to-many match-

ing. In fact, increasing the number of matched nearest neighbours usually increases bias

in exchange for a reduction in variance, and thus may result in a higher number of sta-

tistically significant ATT parameter estimates (possibly incorrectly). We have obtained

similar results by increasing the number of nearest neighbours up to five; however, we

do not present these results here due to space limitations.

6.2 Mahalanobis Matching: All Science Cities

A well-known critique of propensity score matching (King et al., 2011; King and Nielsen,

2016) is that while it may be effective at balancing the treated and matched control sam-

ples, actual pairs of observations with similar p-scores might be very different in their

covariates, biasing the resulting estimates. For this reason, it may be preferable to per-

form Mahalanobis matching or some variation of exact matching instead. On the other

hand, these procedures usually have stronger data requirements. In the context of this

paper, the debate acquires a geographical dimension: while we include polynomials of

municipal coordinates in models for p-score estimation, this solution might not be ef-

fective at matching historically similar places that are also close to each other in space. In

light of this, an alternative solution is to match directly on municipal coordinates, espe-

cially if one believes that relevant unobserved factors with a spatial dimension may be

23The full results of the sensitivity analysis are available upon request.24To provide context, Γ = 2 indicates a simulated unobserved factor that doubles the probability of re-

ceiving the treatment relative to that of not receiving it, or vice versa; such a high value of Γ would berealistic only in presence of very serious threats to our conditional independence assumption (1). Conse-quently, very high “critical” values of Γ∗ associated with a certain outcome – close to 2 or higher – indicatethat the results are likely to be very robust to such threats.

25At a first glance, it may seem counterintuitive that Γ∗ > 1 in the case of outcomes, such as populationin this setting, whose estimated ATT is not statistically different from zero. However, the latter is a conclu-sion derived from a parametric test, while the sensitivity analysis is based on non-parametric Wilcoxonsigned-rank tests. In practice, it is unlikely that the two procedures lead to very divergent conclusions.

23

an issue.

We compare our results obtained via one-to-one PSM with those based on one-to-

one matching on the Mahalanobis distance (2). As we discuss in Section 5, we calculate

the Mahalanobis distance by using the logarithm (and not the levels) of some asymmet-

rically distributed variables; and at the same time we force exact matching on specific

dummy variables (presence of railroads, rivers, lakes) whose corresponding distance

variables are excluded from the calculation of Mahalanobis distance (2). The result-

ing matched samples compare with those obtained via PSM as follows. We are unable

to obtain matches for three Science Cities in total (as opposed to four with PSM) in the

case of non-SME outcomes; while we give up a higher number of them, that is eight in

total (in contrast to two with PSM) in the case of SME outcomes. Because of stronger

data requirements and less flexibility of the method, covariate balance attained via Ma-

halanobis matching compares unfavorably with PSM.26 On the other hand, matched

municipalities are closer to each other, as displayed in Figure 5. The map shows that Ma-

halanobis matching is effective at finding adequately close matches also in areas with a

relatively high concentration of Science Cities, such as the Moscow region and the Urals.

Table 6 displays the empirical results obtained via Mahalanobis matching; estimates

of the ATT for each outcome without and with a bias adjustment term are shown in two

separate columns. The results are very similar to those obtained via PSM; here we high-

light few minor differences. First, the bias adjustment usually implies some reduction

of the point estimates (with the exception of patent outcomes), but no changes in terms

of qualitative conclusions about the statistical significance of the estimated parameters.

The effect on the number of SMEs in manufacturing is no longer statistically significant

in the case of Mahalanobis matching. The results from the sensitivity analysis are also

very similar to those in Table 5: critical values of Γ∗ are now higher for some outcomes

(patents), lower for others (graduate and PhD share, total economy SME labour produc-

tivity), substantially unchanged for the rest; qualitative conclusions are unchanged as

26Covariate balance statistics for Mahalanobis matching are reported in Table E.1 in Appendix E. Theyshow that, particularly for SME outcomes matching, some key covariates (the logs of historical popula-tion, of the number of plants, of municipal area) display a relatively large negative imbalance; that is, themeans for the matched control group are substantially larger in terms of standardised bias. By reasonablyassuming that these outcomes might have a positive effect on today’s outcomes (say, because of urbanpersistence and agglomeration effects in the case of past population, and because of the heritage of pastindustrialisation in the case of the number of plants) we can safely maintain that any bias on the resultingATT estimates due to such imbalance, if present, is negative; implying an interpretation in terms of “lowerbound” effects. In any case, the “bias adjustment” procedure should help mitigate such issues further.

24

the estimated effects that are statistically significant are also very robust. In general, we

interpret the fact that the two different procedures yield substantively similar results as

validation of the empirical strategy adopted in this study.

6.3 Matching Science Cities with no Current Official Status

A different concern about our empirical results is less methodological but more eco-

nomic and institutional in nature. As long as the results are at least in part driven by

Science Cities that have been receiving constant support from the Russian government

even in recent years (that is, some time after the dissolution of the USSR) it would be

difficult to interpret our estimates solely as “long-run” effects, as the results would bear

a mixed interpretation. At the extreme of the spectrum is a scenario in which the results

are only driven by continued funding, although in this case our estimates would still

be useful for a short-run “cost-benefit” analysis of a policy such as the establishment of

Science Cities.

In order to clear these doubts we perform a separate matching analysis which is

largely similar to the one discussed above, with the exception that we now remove Sci-

ence Cities that have the official status of Naukogrady in today’s Russia. As discussed

in section 2, funding for most Science Cities was discontinued after the breakup of the

USSR, and it was resumed only for a subset of them in the early 2000s: primarily the cur-

rent “official” Science Cities. By removing them from the sample, we can assess to a first

degree of approximation to what extent current funding is a factor behind our results.

We replicate both propensity score and Mahalanobis matching on the reduced sample;

in the case of PSM we perform separate propensity score estimation – for both groups

of outcomes – so as to select the models that attain the best covariate balance in the

reduced samples. Tables with covariate balance are provided in Appendix E.27

For brevity, we jump directly to the empirical estimates based on PSM and reported

in Table 7, which we compare directly with those from Table 5. We find the results strik-

ing. While the estimated ATT for the two patent outcomes register point estimates about

60 per cent smaller in both cases (but still statistically significant and robust to unob-

served factors, as given by Γ∗ > 2), all the other outcomes whose effects were previously

found to be positive and statistically significant remain so, with substantially unchanged

point estimates of the ATT parameters. The results based on Mahalanobis matching are

27Matching models are preliminary in both the PSM and Mahalanobis matching cases.

25

presented in Table 8, and display the same pattern. As in the PSM case, all statistically

significant coefficient estimates are very similar to the directly comparable ones in Ta-

ble 6, with the exception of patent outcomes, which are now smaller by 30-50 per cent.

It must be noted, however, that Mahalanobis matching results in a quite small matched

sample for the analysis of SME outcomes.

In short, the results are robust to excluding current official Science Cities from the

sample. The only difference is the reduction in the magnitude of estimated ATT for

patent applications abroad. While the latter finding is intuitive and can be attributed

to the absence of cities with direct government R&D funding, it also reinforces our belief

that the general results for other outcomes – graduate and PhD share, night lights, and

SME labour productivity – are to be interpreted as long-run effects. In fact, if on the con-

trary they were also dependent on current “hidden” funding, one would likely observe

a reduction in point estimates upon excluding current “official” Science Cities from the

sample.28 Still, this leaves open the question of what are the economic mechanisms

driving these results; we turn to this discussion next.

6.4 Mechanisms: Discussion

We summarise the results of our municipal-level analysis as follows. In today’s Rus-

sia, former Soviet-era Science Cities are not more populous than places that were in

most respects similar to them at the end of World War II. However, they host a more

educated population, and they seem to be more economically developed – as demon-

strated by our results on the night lights variable as well as on the two measures of

SME labour productivity. The economic effect of Science Cities occurs on the inten-

sive (rather than extensive) margin, in light of our results on the absolute and relative

numbers of SMEs registered in these municipalities (that is, Science Cities are not much

more “entrepreneurial” than comparable places). In other words, it is a productivity ef-

fect. In terms of innovation, Science Cities apply for more patents abroad, although the

effect is largely driven by the few localities for which the Russian government resumed

funding in the early 2000s.

Having ruled out current funding as a major potential mechanism driving the effects,

28Symmetrically, if the positive albeit smaller effects associated with the patent outcomes in Table 7only depended on the persistence of governmental support for some of today’s non-official Science Cities,and the same mechanism is also hypothesised to affect other outcomes, one would expect higher effectsassociated with the latter in the main results from Table 5.

26

we interpret the results in light of the theoretical contributions cited in the introduction

(Glaeser and Gottlieb, 2009; Kline and Moretti, 2014b). The theory of place-based poli-

cies postulates that long-run effects are only present after the intervention expires as

long as the latter is somehow able to shift economic activity onto a different, “better”

equilibrium. In turn, this is possible if positive agglomeration effects (such as localised

knowledge spillovers) are sufficiently non-linear, as otherwise the prevalence of con-

gestion effects would induce out-migration of workers. In the unique context of Soviet

Science Cities, where the policy specifically targeted R&D personnel, this corresponds

to a scenario in which the policy was initially successful at concentrating human capi-

tal in selected locations; after the suspension of the policy following the dissolution of

the USSR, these human capital differentials persisted and may have translated into a

productivity advantage even through knowledge spillover effects.

This interpretation, however, leaves some questions unanswered while at the same

time raising others. In light of the reduced form evidence, it is at present impossible to

disentangle to what extent the productivity and economic results are a mere “return to

human capital” effect (more educated people make better entrepreneurs), and to what

extent there are any spillover effects (externalities). Looking through the theoretical lens,

the same spatial equilibrium might as well have arisen because of barriers to individual

mobility – unemployed scientists having no other option but to stay where they were and

become entrepreneurs. Existing evidence suggests that this may have been the case: in-

ternal migration rates in Russia in the 1990s were low (see Andrienko and Guriev (2004);

Friebel and Guriev (2005)). Russian anecdotal evidence provides plenty of examples of

both scientists who emigrated and others who became entrepreneurs following the dis-

solution of the USSR, although as discussed in section 2.3, emigration accounted for less

than 3 per cent of the total outflow of staff from the R&D sector. Studying the issue in

more depth is an interesting subject for future work, assuming data exist and can be

accessed.

7 Empirical Results at the Firm Level

In this section we discuss the empirical results at the firm level, assessing the “spatial

reach” of Science Cities as expressed by our hypothesis HF.1. We separately focus on

probit estimates of binary innovation outcomes, and on OLS regressions featuring con-

tinuous measures of firm performance on the left hand side. In addition, we provide a

27

brief discussion of these results in light of those from the municipal-level analysis.

7.1 Binary Innovation Outcomes

Table 9 presents the results from the estimation of probit model (4) for four separate

firm-level outcome binary outcomes I f r : whether a firm engages in any R&D activity,

whether in the three years prior to the survey the firm has introduced either a product

or a process new to the firm or beyond, and finally looking at product and process inno-

vation separately. In addition to our measure of distance of the firm from Science Cities,

on the right-hand side of the model we include a set of controls specified in section 5.2.

We cluster standard errors at the municipal level.

The estimates of our coefficient of interest β are as follows: negative and not statis-

tically different from zero for engaging in R&D, positive and statistically different from

zero (at the 5 per cent significance level) for engaging in either product or process in-

novation (or both), and finally positive but not statistically significant from zero for en-

gaging in product and process innovation separately.29 In short, there is mixed evidence

of a spatial correlation between the location of science cities and firm-level innovation:

while firms closer to a Science City are more likely to innovate, the β coefficients for the

two “disaggregated” outcomes from columns (3) and (4) of the table are not precisely

estimated.

While it is possible to find probabilistic explanations of these results favourable to

our initial hypothesis – say, statistical tests of this relationships may have too low of a

power if one restricts the analysis to specific types of innovation – one could also see the

significant β parameter from column (2) itself as a mere random occurrence. Taken at

face value, however, that estimated coefficient is economically significant. Back-of-the-

envelope calculations of the marginal effect from column (2) suggest that if the average

firm were to move from an area with no Science Cities to a Science City, the likelihood

that it would engage in (product or process) innovation would increase by 17.7 per cent

29As for the other explanatory variables for firm innovation, Table 9 indicates that larger and older firmsare more likely to innovate (perhaps because newer firms start closer to the technological frontier), andso are firms in manufacturing and the ones employing a relatively more skilled labour force. While firmswho mostly trade on the national market are more likely to conduct R&D and introduce new productsand processes, those whose market is more local are more likely to engage in process innovation. Firmswhose main market is international, are more likely to invest in R&D and engage in product innovation;being controlled by the state seems to reduce the likelihood to engage in innovation; being controlled byforeign owners, on the other hand, seems to have no effect at all.

28

(with 3.6 per cent p-value).30 This effect may be driven by firms that engage in product

and process innovation simultaneously, rather than separately in either of the two. This

conclusion, however, implicitly assumes that there are no unobserved determinants of

innovation that correlate with the observed location of cities and that are not appropri-

ately accounted for by our set of controls.

7.2 Other Continuous Outcomes

Table 10 reports estimates of model (5) for three firm-level outcomes: log operating rev-

enue (also controlling for conventional production inputs), log labour productivity, and

the share of firm sales that are due to innovative products. As above, we include addi-

tional controls, region fixed effects, and cluster standard errors at the municipal level. In

all three regressions, the estimated β coefficient is negative and not statistically differ-

ent from zero; that is, there is no association between the performance of firms and their

proximity to Science Cities. These more negative results could be reconciled with those

relative to firm-level innovation by postulating a weak relationship between innovation

and performance of Russian firms, possibly due to institutional barriers hampering the

growth of more dynamic Russian enterprises.

7.3 Results: Discussion

Our firm-level results indicate that there is mixed evidence of a spatial relationship be-

tween the location of Science Cities and Russian firms most inclined to engage in inno-

vation, coupled with a decidedly unfavourable evidence for the existence of an analo-

gous link with measures of firm performance. In light of the results from our municipal-

level analysis, we argue that the scope of Science Cities’ effects on innovation and eco-

nomic outcomes are limited in space, which is consistent with fast distance decay of

knowledge spillovers as evidenced in prior studies, such as (Lychagin et al., 2016). They

may also be specific to certain sectors within that limited space, rather than benefitting

firms in any sector. More research is certainly needed for a better understanding of the

spatial extent of knowledge flows, with particular reference to the Russian context.

30We estimate this as the change in the likelihood in engaging in some kind of innovation that is asso-ciated with increasing the main regressor

∑Ss=1 exp

[−Dist(

f , s)]

in (4) from 0 to 1, and we calculate theassociated standard error via a Delta Method approximation. Both this figure and its attached interpreta-tion depend on the functional form assumption of the probit model (4).

29

8 Conclusion

In this paper we have analysed the long-run effects of a historical placed-based policy:

the creation of R&D-focused Science Cities in Soviet Russia. Both the initial establish-

ment and the eventual partial suspension of this program was largely guided by political

factors that are arguably exogenous to drivers of current social and economic conditions

of Russian cities. We compare municipalities with Science Cities to other municipalities

that were observationally similar to them at the time of their selection, and we compute

differences in the current characteristics between the two groups. We find that the mu-

nicipalities with former Science Cities, while not significantly larger, still host a more

educated population today, apply for more patents abroad, and host more productive

SMEs (although not a higher number of them). With a separate firm-level analysis, how-

ever, we can attest only limited evidence in support of the hypothesis that the effect of

Science Cities extends beyond the municipal borders.

Because our results remain unchanged even after excluding municipalities with Sci-

ence Cities that still receive support by the Russian government from the sample, we

conjecture that they are primarily driven by the local persistence of a more educated

population, an institutional heritage from the Soviet era. With the information currently

at our disposal, we are unable to distinguish whether this is mostly caused by barriers to

labour mobility, or by localised knowledge spillovers that would, on one hand, change

the spatial equilibrium and, on the other hand, contribute directly to the observed ef-

fects, though there is some evidence suggesting there were barriers to labour mobility.

We leave the exploration of this problem to future work. In either case, we believe that

ours is a valuable contribution to the existing literature on place-based policies, which

up to now has found only limited evidence in favour of long-run effects following the

suspension of a program. More generally, our results are also informative for science

and innovation policy, both in the context of emerging economies such as Russia as well

as developed economies.

References

Abadie, Alberto and Guido W. Imbens (2006) “Large Sample Properties of Matching Es-

timators for Average Treatment Effects,” Econometrica, Vol. 74, No. 1, pp. 235–267.

30

(2011) “Bias-corrected matching estimators for average treatment effects,” Jour-

nal of Business and Economic Statistics, Vol. 29, No. 1, pp. 1–11.

Aguirrechu, Alexander A. (2009) Russian science cities: History of formation and develop-

ment (Naukogradi Rossii: Istoriya formirovaniya i razvitiya), Moscow: Moscow Uni-

versity Press. Available in Russian only.

Ambler, John, Denis J. B. Shaw, and Leslie Symons eds. (1985) Soviet and East European

Transport Problems, London: Croom Helm.

Andrienko, Yuri and Sergei Guriev (2004) “Determinants of interregional mobility in

Russia: Evidence from panel data,” Economics of Transition, Vol. 12, No. 1, pp. 1–27.

Audretsch, David B. and Maryann P. Feldman (1996) “R&D Spillovers and the Geography

of Innovation and Production,” American Economic Review, Vol. 86, No. 3, pp. 630–

640.

Beaudry, Catherine and Andrea Schiffauerova (2009) “Who’s right, Marshall or Jacobs?

The localization versus urbanization debate,” Research Policy, Vol. 38, No. 2, pp. 318–

337.

Bloom, Nicholas, Mark Schankerman, and John Van Reenen (2013) “Identifying Tech-

nology Spillovers and Product Market Rivalry,” Econometrica, Vol. 81, No. 4, pp. 1347–

1393.

Boschma, Ron and Koen Frenken (2011) “The emerging empirics of evolutionary eco-

nomic geography,” Journal of Economic Geography, Vol. 11, No. 2, pp. 295–307.

Breschi, Stefano and Francesco Lissoni (2009) “Mobility of skilled workers and co-

invention networks: an anatomy of localized knowledge flows,” Journal of Economic

Geography, Vol. 9, No. 4, pp. 439–468.

Bronzini, Raffaello and Guido de Blasio (2006) “Evaluating the impact of investment in-

centives: The case of Italy’s Law 488/1992,” Journal of Urban Economics, Vol. 60, No.

2, pp. 327–349.

Busso, Matias, Jesse Gregory, and Patrick M. Kline (2013) “Assessing the incidence and

efficiency of a prominent place based policy,” American Economic Review, Vol. 103,

No. 2, pp. 897–951.

31

Central management unit of the military communications of the Red Army (1943)

Schemes of railways and waterways of the USSR (Shemi zheleznih dorog i vodnih putey

soobsheniya SSSR): Military Publishing of the People’s Commisariat of Defence. Avail-

able in Russian only at http://istmat.info/files/uploads/45009/shemy_zhd_

i_vodnyh_putey_sssr_1943.pdf (last accessed on 30 November 2016).

Cheremukhin, Anton, Mikhail Golosov, Sergei Guriev, and Aleh Tsyvinski (2017) “The

Industrialization and Economic Development of Russia through the Lens of a Neo-

classical Growth Model,” Review of Economic Studies. Forthcoming.

Commander, Simon, Zlatko Nikoloski, and Alexander Plekhanov (2011) “Employment

concentration and resource allocation: one-company towns in Russia,” Working Pa-

per 130, EBRD.

Cooper, Julian M. (2012) “Science-technology policy and innovation in the USSR,”

slides, CEELBAS Workshop "Russia’s Skolkovo in Comparative and Historical Perspec-

tive", held on 12 June 2012 at UCL SSEES. Available at http://www.ceelbas.ac.uk/

workshops/skolkovo/Cooper (last accessed on 30 November 2016).

Criscuolo, Chiara, Ralf Martin, Henry Overman, and John Van Reenen (2012) “The

Causal Effects of an Industrial Policy,” NBER Working Paper 17842, National Bureau

of Economic Research.

Dexter, Keith and Ivan Rodionov (2016) “The Factories, Research and Design Estab-

lishments of the Soviet Defence Industry: a Guide. Version 17,” working paper,

The University of Warwick, Department of Economics. Available at http://www2.

warwick.ac.uk/fac/soc/economics/staff/mharrison/vpk/data/ (last accessed

on 29 November 2016).

Donaldson, Dave and Adam Storeygard (2016) “Thew view from above: Applications of

satellite data in economics,” Journal of Economic Perspectives, Vol. 30, No. 4, pp. 171–

198.

Duranton, Gilles and Diego Puga (2004) “Micro-Foundations of Urban Agglomeration

Economies,” in J. Vernon Henderson and Jacques-François Thisse eds. Handbook of

Regional and Urban Economics, Vol. 4: Cities and Geography, Chap. 48, pp. 2063–2117.

32

Ellison, Glenn, Edward L. Glaeser, and William R. Kerr (2010) “What causes industry ag-

glomeration? Evidence from coagglomeration patterns,” American Economic Review,

Vol. 100, No. 3, pp. 1195–1213.

Friebel, Guido and Sergei Guriev (2005) “Attaching workers through in-kind payments:

Theory and evidence from Russia,” World Bank Economic Review, Vol. 19, No. 2, pp.

175–202.

Ganguli, Ina (2014) “Immigration & ideas: What did Russian scientists ’bring’ to the US?”

Working Paper 30, Stockholm Institute of Transition Economics.

Givord, Pauline, Rolande Rathelot, and Patrick Sillard (2013) “Place-based tax exemp-

tions and displacement effects: An evaluation of the Zones Franches Urbaines pro-

gram,” Regional Science and Urban Economics, Vol. 43, No. 1, pp. 151–163.

Glaeser, Edward L., Hedi D. Kallal, José A. Scheinkman, and Andrei Schleifer (1992)

“Growth in Cities,” Journal of Political Economy, Vol. 100, No. 6, pp. 1126–1152.

Glaeser, Edward L. and Joshua D. Gottlieb (2008) “The economics of place-making poli-

cies,” Brookings Papers on Economic Activity, pp. 155–239, Spring.

(2009) “The Wealth of Cities: Agglomeration Economies and Spatial Equilibrium

in the United States,” Journal of Economic Literature, Vol. 74, No. 4, pp. 983–1028.

Gokhberg, Leonid (1997) “Transformation of the Soviet R&D System,” in Leonid

Gokhberg, Merton J. Peck, and János Gács eds. Russian Applied Research and Devel-

opment: Its Problems and its Promise, Laxenburg, Austria: International Institute for

Applied Systems Analysis, pp. 9–33. Available at http://pure.iiasa.ac.at/5156/

1/RR-97-007.pdf.

Gokhberg, Leonid and Levan Mindely (1993) “Soviet R&D resources: Basic characteris-

tics,” in Serguei Glaziev and Christoph M. Schneider eds. Research and Development

Management in the Transition to a Market Economy, Laxenburg, Austria: IIASA, pp. 3–

23. Available at http://pure.iiasa.ac.at/3813/1/CP-93-001.pdf (last accessed

on 30 March 2017).

Gokhberg, Leonid and Elena Nekipelova (2002) “International migration of scientists

and engineers in Russia,” in International mobility of the highly skilled, Paris: OECD,

pp. 177–187.

33

Graham, Loren and Irina Dezhina (2008) Science in the new Russia: Crisis, aid, reform,

Bloomington, IN: Indiana University Press.

Greenstone, Michael, Richard Hornbeck, and Enrico Moretti (2010) “Identifying agglom-

eration spillovers: Evidence from winners and losers of large plant openings,” Journal

of Political Economy, Vol. 118, No. 3, pp. 536–598.

Gu, Xing Sam and Paul R. Rosenbaum (1993) “Comparison of Multivariate Match-

ing Methods: Structures, Distances, and Algorithms,” Journal of Computational and

Graphical Statistics, Vol. 2, No. 4, pp. 405–420.

Hodler, Roland and Paul A. Raschky (2014) “Regional favoritism,” The Quarterly Journal

of Economics, Vol. 129, No. 2, pp. 995–1033.

Ivanov, Denis S. (2016) “Human Capital and Knowledge-Intensive Industries Location:

Evidence from Soviet Legacy in Russia,” Journal of Economic History, Vol. 76, No. 3,

pp. 736–768.

Jaffe, Adam B. (1989) “Real Effects of Academic Research,” American Economic Review,

Vol. 79, No. 5, pp. 957–970.

Jaffe, Adam B., Manual Trajtenberg, and Rebecca Henderson (1993) “Geographic local-

ization of knowledge spillovers as evidenced by patent citations,” The Quarterly Jour-

nal of Economics, Vol. 108, No. 3, pp. 577–598.

King, Gary, Richard Nielsen, Carter Coberley, James E. Pope, and Aaron Wells (2011)

“Comparative Effectiveness of Matching Methods for Causal Inference,” working pa-

per. Available at http://http://gking.harvard.edu/files/psparadox.pdf (last

accessed on 29 March 2017).

King, Gary and Richard Nielsen (2016) “Why propensity scores should not be used for

matching,” working paper. Available at http://gking.harvard.edu/files/gking/

files/psnot.pdf (last accessed on 29 March 2017).

Kline, Patrick M. and Enrico Moretti (2014a) “Local Economic Development, Agglomer-

ation Economies, and the Big Push: 100 Years of Evidence from the Tennessee Valley

Authority,” The Quarterly Journal of Economics, Vol. 121, No. 1, pp. 275–331.

34

(2014b) “People, Places, and Public Policy: Some Simple Welfare Economics of

Local Economic Development Programs,” Annual Review of Economics, Vol. 6, pp.

629–662.

Lappo, G. M. and P. M. Polyan (2008) “Naukograds of Russia: Yesterday’s forbidden and

semi-forbidden cities - today’s growth points (Naukogrady Rossii: Vcherashniye za-

pretnyye i poluzapretnyye goroda - segodnyashniye tochki rosta),” World of Russia

(Mir Rossii), No. 1, pp. 20–49. Available in Russian only.

Lychagin, Sergey, Joris Pinkse, Margaret E. Slade, and John Van Reenen (2016) “Spillovers

in space: Does geography matter?” Journal of Industrial Economics, Vol. 64, No. 2, pp.

295–335.

Marshall, Alfred (1890) Principles of economics, London: Macmillan, 8th edition.

Mikhailova, Tatiana (2012) “Where Russians Should Live: a Counterfactual Alternative

to Soviet Location Policy,” in MPRA Paper 35938: University Library of Munich, Ger-

many.

Moretti, Enrico (2004) “Workers’ education, spillovers and productivity: Evidence from

plant-level production functions,” American Economic Review, Vol. 94, No. 3, pp. 656–

690.

Moretti, Enrico, Claudia Steinwender, and John Van Reenen (2016) “The Intellectual

Spoils of War? Defense R&D, Productivity and Spillovers.” mimeo.

NAS (2002) Successes and difficulties of small innovative firms in Russian nuclear cities:

Proceedings of a Russian-American workshop. Committee on Small Innovative Firms

in Russian Nuclear Cities, Office for Central Europe and Eurasia Development, Se-

curity, and Cooperation, National Research Council, in cooperation with the Insti-

tute of Physics and Power Engineering, Obninsk, Russia, Washington, D.C.: National

Academy of Sciences. Available at http://www.nap.edu/catalog/10392.html (last

accessed on 29 September 2016).

Neumark, David and Jed Kolko (2010) “Do enterprise zones create jobs? Evidence from

California’s Enterprise Zone Program,” Journal of Urban Economics, Vol. 68, No. 1, pp.

1–19.

35

Neumark, David and Helen Simpson (2014) “Place-Based Policies,” NBER Working Paper

20049, National Bureau of Economic Research.

OECD and Statistical Office of the European Communities (2005) Oslo Manual: Guide-

lines for Collecting and Interpreting Innovation Data, 3rd Edition, The measurement

of Scientific and Technological Activities, Luxembourg: OECD Publishing, pp.162.

Rosenbaum, Paul R. (2002) Design of Observational Studies: Springer.

Rubin, Donald B. and Neal Thomas (1996) “Matching Using Estimated Propensity

Scores: Relating Theory to Practice,” Biometrics, Vol. 52, No. 1, pp. 249–264.

Ruchnov, K. M. and E. G. Zaitseva (2011) Crisis of Russian science cities (Krizis

naukogradov Rossii). Available in Russian only at http://www.mosveo.ru/images/

stories/00008.doc (last accessed on 20 March 2013).

Saltykov, Boris G. (1997) “The reform of Russian science,” Nature, Vol. 388, No. 6637, pp.

16–18.

Schneider, Christoph M. (1994) Research and development management: From the Soviet

Union to Russia, Heidelberg: Physica-Verlag. Available at http://pure.iiasa.ac.

at/3993/1/XB-94-002.pdf (last accessed on 30 March 2017).

Storeygard, Adam (2016) “Farther on down the road: Transport costs, trade and urban

growth in Sub-Saharan Africa,” Review of Economic Studies, Vol. 83, No. 3, pp. 1263–

1295.

Thompson, Peter and Melanie Fox-Kean (2005) “Patent Citations and the Geography of

Knowledge Spillovers: A Reassessment,” The American Economic Review, Vol. 95, No.

1, pp. 450–460.

Vernadsky State Geological Museum and U.S. Geological Survey, 20010600 (2001)

“rails.shp. Railroads of the Former Soviet Union,” U.S. Geological Survey, Denver,

CO. Available at http://pubs.usgs.gov/of/2001/ofr-01-104/fsucoal/html/

data1.htm\#rail (last accessed on 29 November 2016).

Wang, Jin (2013) “The economic impact of Special Economic Zones: Evidence from Chi-

nese municipalities,” Journal of Development Economics, Vol. 101, No. 2, pp. 133–147,

March.

36

Zhao, Zhong (2004) “Using Matching to Estimate Treatment Effects: Data Requirements,

Matching Metrics, and Monte Carlo Evidence,” The Review of Economics and Statistics,

Vol. 86, No. 1, pp. 91–107.

37

Figures

Figure 1: Location of science cities and regions covered in BEEPS V Russia

Source: Table A.1 and BEEPS V Russia.

Figure 2: Gross expenditures on R&D (GERD) in Russia, 1989-2010

Source: Gokhberg (1997), Russian Statistical Yearbooks (various years) and OECD MainScience and Technology Indicators (MSTI) database.

38

Figure 3: Researchers in Russia, 1989-2010

Source: Gokhberg (1997), Russian Statistical Yearbooks (various years) and OECD MainScience and Technology Indicators (MSTI) database.

Figure 4: Defense R&D in Russia and elsewhere, 1994-2003

Source: Gokhberg (1997) and OECD Main Science and Technology Indicators (MSTI)database.

39

Figure 5: Science Cities and their matches

(a) Propensity score matching

(b) Mahalanobis matching

Source: Authors’ calculations.

40

Tables

Table 1: Municipal-level data: Descriptive statistics

Municipalities Municipalitieswith Science Cities without Science CitiesObs. Mean (SE) Obs. Mean (SE) p-value

Latitude 88 55.664 2250 53.981 0.000(0.391) (0.108)

Longitude 88 49.771 2250 59.955 0.000(2.387) (0.620)

Average temperature in January (°C) 88 -11.632 2250 -13.559 0.000(0.410) (0.149)

Average temperature in July (°C) 88 18.535 2250 18.755 0.247(0.181) (0.056)

Average altitude 88 0.169 2250 0.267 0.000(0.010) (0.007)

Minimum distance from railroad 88 0.007 2250 0.078 0.000(0.001) (0.005)

Minimum distance from river 88 0.032 2250 0.056 0.000(0.004) (0.001)

Minimum distance from lake 88 0.118 2250 0.172 0.000(0.009) (0.003)

Minimum distance from coast 88 0.725 2250 0.730 0.917(0.044) (0.010)

Minimum distance from USSR border 88 0.665 2250 0.679 0.723(0.037) (0.009)

Population in 1959 88 67.583 2250 49.573 0.167(12.516) (3.242)

Number of universities in 1959 88 0.557 2250 0.196 0.132(0.224) (0.046)

Number of plants in 1947 88 6.205 2250 2.484 0.023(1.458) (0.697)

Number of R&D institutes in 1959 88 0.807 2250 0.412 0.242(0.253) (0.222)

Area in km2 88 0.692 2250 7.108 0.000(0.116) (0.637)

Patent applications, 2006-2015, simple 88 37.875 2250 6.025 0.000(8.033) (3.375)

Patent applications, 2006-2015, fractional 88 13.909 2250 2.265 0.002(3.489) (1.210)

Night lights, 2009-2011 88 30.611 2250 7.638 0.000(2.124) (0.272)

Population in 2010 88 131.557 2250 58.324 0.001(21.169) (5.871)

Graduate share in 2010 88 0.225 2250 0.110 0.000(0.008) (0.001)

PhD share in 2010 88 0.006 2250 0.003 0.000(0.000) (0.000)

Number of SMEs in 2010 (thousands, all) 69 3239.725 2140 1189.833 0.008(742.669) (67.367)

SMEs per 1000 people (all) 69 0.025 2159 0.027 0.086(0.001) (0.000)

SME labor productivity (all) 69 1643.995 2153 794.105 0.000(84.513) (9.213)

Number of SMEs in 2010 (thousands, manufacturing) 69 395.073 2038 119.546 0.010(103.133) (7.535)

SMEs per 1000 people (manufacturing) 69 0.002 2038 0.002 0.066(0.000) (0.000)

SME labour productivity (manufacturing) 67 1438.443 2014 768.462 0.000(84.554) (20.805)

41

Table 2: Firm-level data: Descriptive statistics

Obs Mean Std. dev. Min MaxYoung firms (0-5 years) 4220 0.297 0.457 0 1

50% or more foreign owned 4220 0.026 0.158 0 1

50% or more state owned 4220 0.012 0.108 0 1

Direct exporter 4220 0.100 0.300 0 1

Main market: local 4220 0.697 0.460 0 1

Main market: national 4220 0.288 0.453 0 1

% of employees with a completed university degree 4045 52.505 30.521 0 100

Manufacturing 4220 0.325 0.469 0 1

Located in a city with population over 1 million 4220 0.255 0.436 0 1

Minimum distance of a firm from a Science City (in km) 4220 154.410 237.742 0.198 1358.035

Exponential 1 decay distance of a firm from a Science City 4220 0.012 0.067 0.000 0.822

Log (number of employees), Orbis 2979 3.750 0.860 0.693 8.860

Log (capital), Orbis 3027 5.603 2.089 -2.659 12.916

Log (materials), Orbis 2936 6.212 1.964 -3.912 12.839

Log (operating revenue), Orbis 2980 6.442 1.862 -1.966 13.036

Total factor productivity 2979 0.079 1.356 -8.194 6.544

Log labour productivity, Orbis 2979 2.690 1.343 -5.577 9.167

R&D (dummy) 4220 0.111 0.314 0 1

% of innovative products in sales 4101 3.855 13.937 0 100

Technological innovation (dummy) 4220 0.216 0.411 0 1

Product innovation (dummy) 4220 0.129 0.335 0 1

Process innovation (dummy) 4220 0.141 0.348 0 1

Source: Authors’ calculations based on BEEPS V and Bureau Van Dijk’s Orbis.

42

Table 3: Propensity Score Models (Probit Estimates)

(1): Model for (2): Model forNon-SME outcomes SME outcomes

Latitude 1.105∗∗∗ (0.409) 0.225∗∗ (0.110)Latitude (Sq.) -0.007∗∗ (0.003)Longitude 0.083 (0.053) -0.056 (0.091)Longitude (Sq.) 0.000 (0.000)Latitude × Longitude -0.001 (0.001) 0.001 (0.001)Average temperature in January (°C) 0.246∗∗∗ (0.070) 0.098 (0.119)Average temperature in January (°C) (Sq.) -0.004 (0.004)Average temperature in July (°C) 0.203∗ (0.114) 0.737 (0.475)Average temperature in July (°C) (Sq.) -0.013 (0.011)Average altitude 2.434∗∗ (0.983) 3.139∗ (1.783)Average altitude (Sq.) -1.278 (1.931)Minimum distance from railroad -3.981 (14.767) -7.873 (14.053)Minimum distance from railroad (Sq.) -273.8 (326.7) -195.2 (303.4)Minimum distance from river -3.279∗ (1.826) -7.947∗∗ (3.610)Minimum distance from river (Sq.) 31.481 (21.661)Minimum distance from lake 1.592 (1.915) 2.048 (1.899)Minimum distance from lake (Sq.) -7.449 (5.281) -9.541∗ (5.271)Minimum distance from coast -0.739∗ (0.396) -0.267 (0.404)Minimum distance from USSR border 1.285 (0.797) 0.162 (0.351)Minimum distance from USSR border (Sq.) -0.660 (0.458)Population in 1959 -0.003∗ (0.002) -0.003∗ (0.002)Number of universities in 1959 -0.231 (0.155) 0.001 (0.085)Number of universities in 1959 (Sq.) 0.029∗∗ (0.014)Number of plants in 1947 0.046∗∗ (0.021) 0.016 (0.014)Number of plants in 1947 (Sq.) -0.001∗∗ (0.000)Number of R&D institutes in 1959 0.146∗ (0.080) 0.208∗∗∗ (0.079)Number of R&D institutes in 1959 (Sq.) -0.002 (0.003) -0.005∗ (0.003)Area in km2 -0.342∗∗∗ (0.063) -0.318∗∗∗ (0.063)Area in km2 (Sq.) 0.001∗ (0.001)Constant -43.729∗∗∗ (14.945) -21.644∗∗ (9.765)Log-Likelihood -264.23 -268.87Pseudo R2 0.2953 0.2829Model χ2 (p-value) 221.42 (0.0000) 212.13 (0.0000)Observations 2338 2338

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is the p-valueassociated to each parameter estimate (standard errors are reported in parentheses). The fol-lowing variables, along with the associated squared terms (if present), enter the model multi-plied by a scale factor of 10−3: average altitude; minimum distances from any railroad, river,lake, coast, the USSR border; population in 1959; area in km2.

43

Table 4: Covariate Balance: Propensity Score Matching, all Science Cities

Matched Sample: non-SME OutcomesVariable Treated Control Std. bias t-stat. p-value V-ratioLatitude 55.664 55.871 -5.4 -0.36 0.722 0.94Longitude 49.771 49.861 -0.4 -0.03 0.979 0.92Average temperature in January (°C) -11.632 -11.706 1.8 0.12 0.905 0.83Average temperature in July (°C) 18.535 18.424 6.1 0.40 0.689 0.90Average altitude 0.1686 0.1706 -1.8 -0.12 0.908 0.63Minimum distance from railroad 0.0068 0.0074 -5.8 -0.38 0.707 1.08Minimum distance from river 0.0319 0.0312 2.2 0.14 0.886 1.39Minimum distance from lake 0.1184 0.1185 -0.1 -0.01 0.995 0.94Minimum distance from coast 0.7246 0.7221 0.6 0.04 0.969 0.86Minimum distance from USSR border 0.6651 0.6404 6.9 0.45 0.650 0.70Population in 1959 67.583 65.933 1.5 0.10 0.922 1.44Number of universities in 1959 0.5568 0.5238 1.7 0.11 0.914 1.31Number of plants in 1947 6.2045 5.9286 2.1 0.14 0.889 1.34Number of R&D Institutes in 1959 0.8068 0.6786 5.5 0.36 0.719 1.31Area in km2 0.6923 0.7791 -8.6 -0.56 0.574 1.08

Matched Sample: SME OutcomesVariable Treated Control Std. bias t-stat. p-value V-ratioLatitude 55.664 56.027 -9.6 -0.59 0.555 0.95Longitude 49.771 48.819 4.3 0.26 0.792 0.99Average temperature in January (°C) -11.632 -11.714 2.1 0.13 0.896 1.01Average temperature in July (°C) 18.535 18.445 5.3 0.32 0.746 1.04Average altitude 0.1686 0.1666 1.9 0.12 0.904 0.78Minimum distance from railroad 0.0068 0.0064 5.3 0.32 0.747 1.63Minimum distance from river 0.0319 0.0325 -1.9 -0.12 0.906 1.46Minimum distance from lake 0.1184 0.1221 -4.0 -0.25 0.803 0.81Minimum distance from coast 0.7246 0.7124 3.1 0.19 0.850 1.15Minimum distance from USSR border 0.6651 0.6829 -5.0 -0.31 0.758 0.94Population in 1959 67.583 76.074 -7.4 -0.46 0.650 1.65Number of universities in 1959 0.5568 0.5970 -2.1 -0.13 0.898 1.82Number of plants in 1947 6.2045 6.8806 -5.1 -0.31 0.754 1.50Number of R&D institutes in 1959 0.8068 0.7761 1.4 0.09 0.930 1.63Area in km2 0.6923 0.7183 -2.6 -0.16 0.875 1.32

Notes: For each variable, the table reports its mean for both the treated and control groups of thematched sample in question, along with the corresponding standardised bias (the treated-control dif-ference as a percentage of the combined standard deviation of the matched sample). In addition, foreach variable the table reports the t-statistic and the associated p-value for the difference in the twomeans, as well as the ratio between the variances of the treated and the variance of the control groupfor the variable in question. Both matched samples are obtained via 1:1 nearest neighbour matching onthe estimated propensity score (with replacement).

44

Table 5: Municipal-level Analysis: Propensity Score Matching, all Science Cities

Whole sample Matched sample (1 nearest neighbour)

Outcome Raw difference T C ATT Γ∗ (α= .05)

Patent applications, simple31.850***

84 7528.167***

2.80(8.675) (7.352)

Patent applications, fractional11.644***

84 759.953***

2.55(3.676) (3.336)

Night lights, 2009-201122.973***

84 757.587***

1.55(2.130) (2.271)

Population in 195973.233***

84 7532.164

1.20(21.861) (25.861)

Graduate share in 20100.115***

84 750.074***

5.55(0.008) (0.008)

PhD share in 20100.003***

84 750.003***

2.75(0.000) (0.000)

Number of SMEs in 2010 (thousands, all)2.050***

67 56-0.907

1.00(0.740) (0.827)

Number of SMEs in 2010 (thousands, manufacturing)0.275***

67 56-0.254*

1.35(0.103) (0.131)

SMEs per 1000 people (all)-2.042*

67 56-4.372***

1.20(1.222) (1.557)

SMEs per 1000 people (manufacturing)0.240*

67 560.077

1.00(0.128) (0.130)

SME labour productivity (all)0.850***

67 560.526***

4.10(0.084) (0.068)

SME labour productivity (manufacturing)0.670***

67 560.486***

1.65(0.086) (0.105)

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is the p-value associated to each param-eter estimate (standard errors are reported in parentheses). In the matched sample, T is the number of matched treatedobservations; C is the number of matched controls; ‘ATT’ is the estimate of (3), where standard errors are computed fol-lowing Abadie and Imbens (2006) and account for the fact that the propensity score is estimated; while Γ∗ is the minimumvalue of parameter Γ ≥ 1, selected from a grid spaced by intervals of 0.05 length, such that in a sensitivity analysis à laRosenbaum (2002) the set of Wilcoxon signed-rank tests associated with Γ∗ do not simultaneously reject the null hypoth-esis that the outcome variable is not different across the treated and control samples, for tests with α= 0.05 type I error. Ahigher value of Γ is associated to a stronger simulated unobserved factor which affects both the outcome and the proba-bility of receiving the treatment. Complete results of the sensitivity analysis for specific outcomes and for different valuesof Γ are available upon request.

45

Table 6: Municipal-level Analysis: Mahalanobis Matching, all Science Cities

Whole sample Matched sample (1 nearest neighbour)

Outcome Raw difference T C ATT ATT b.a. Γ∗ (α= .05)

Patent applications, simple31.850***

85 7029.835*** 28.556***

6.25(8.675) (7.345) (7.215)

Patent applications, fractional11.644***

85 7011.138*** 10.672***

5.55(3.676) (3.352) (3.330)

Night lights22.973***

85 707.234*** 5.512***

1.60(2.130) (2.147) (2.034)

Population in 201073.233***

85 7024.486* 18.891

1.90(21.861) (13.157) (11.860)

Graduate share in 20100.115***

85 700.064*** 0.055***

3.95(0.008) (0.012) (0.011)

PhD share in 20100.003***

85 700.002*** 0.002***

2.05(0.000) (0.000) (0.000)

Number of SMEs (thousands, all)2.050***

61 520.238 0.230

1.20(0.740) (0.349) (0.362)

Number of SMEs (thousands, manufacturing)0.275***

61 520.077 0.062

1.20(0.103) (0.069) (0.064)

SMEs per 1000 people (all)-2.042*

61 52-2.046 -2.975*

1.00(1.222) (1.777) (1.841)

SMEs per 1000 people (manufacturing)0.240*

61 52-0.181 -0.303

1.00(0.128) (0.230) (0.237)

SME labour productivity (All)0.850***

61 520.420*** 0.372***

2.85(0.084) (0.085) (0.085)

SME labour Productivity (manufacturing)0.670***

61 520.311** 0.253**

1.65(0.086) (0.122) (0.119)

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is the p-value associated to each parameterestimate (standard errors are reported in parentheses). In the matched sample, T is the number of matched treated observa-tions; C is the number of matched controls; ‘ATT’ and ‘ATT b.a.’ are two estimates of (3) respectively excluding and includinga bias-adjustment term as per Abadie and Imbens (2011) (in both cases, standard errors are computed following Abadie andImbens (2006)); while Γ∗ is the minimum value of parameter Γ ≥ 1, selected from a grid spaced by intervals of 0.05 length,such that in a sensitivity analysis à la Rosenbaum (2002) the set of Wilcoxon signed-rank tests associated with Γ∗ do not si-multaneously reject the null hypothesis that the outcome variable is not different across the treated and control samples, fortests with α = .05 type I error. A higher value of Γ is associated to a stronger simulated unobserved factor which affects boththe outcome and the probability of receiving the treatment. Complete results of the sensitivity analysis for specific outcomesand for different values of Γ are available upon request.

46

Table 7: Municipal-level Analysis: Propensity Score Matching, excluding current Science Cities

Whole sample Matched sample (1 nearest neighbour)

Outcome Raw difference T C ATT Γ∗ (α= .05)

Patent applications, simple23.448***

70 6313.729***

2.65(8.978) (3.209)

Patent applications, fractional7.254***

70 634.067***

2.35(2.703) (0.811)

Night lights20.101***

70 635.870**

1.25(2.318) (2.597)

Population in 201082.854***

70 6325.783

1.50(25.398) (24.905)

Graduate share in 20100.103***

70 630.063***

3.05(0.009) (0.010)

PhD share in 20100.003***

70 630.002***

2.15(0.000) (0.000)

Number of SMEs (thousands, all)2.356***

54 490.576

1.15(0.900) (0.800)

Number of SMEs (thousands, manufacturing)0.314**

54 490.109

1.45(0.125) (0.131)

SMEs per 1000 people (all)-2.639*

54 49-1.538

1.15(1.347) (1.848)

SMEs per 1000 people (manufacturing)0.095

54 49-0.101

1.00(0.128) (0.233)

SME labour productivity (all)0.799***

54 490.467***

2.45(0.097) (0.104)

SME labour productivity (manufacturing)0.636***

54 490.506***

1.35(0.100) (0.109)

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is the p-value associated toeach parameter estimate (standard errors are reported in parentheses). In the matched sample, T is the numberof matched treated observations; C is the number of matched controls; ‘ATT’ is the estimate of (3), where standarderrors are computed following Abadie and Imbens (2006) and account for the fact that the propensity score is esti-mated; while Γ∗ is the minimum value of parameter Γ≥ 1, selected from a grid spaced by intervals of 0.05 length,such that in a sensitivity analysis à la Rosenbaum (2002) the set of Wilcoxon signed-rank tests associated with Γ∗do not simultaneously reject the null hypothesis that the outcome variable is not different across the treated andcontrol samples, for tests with α= 0.05 type I error. A higher value of Γ is associated to a stronger simulated unob-served factor which affects both the outcome and the probability of receiving the treatment. Complete results ofthe sensitivity analysis for specific outcomes and for different values of Γ are available upon request.

47

Table 8: Municipal-level Analysis: Mahalanobis Matching, excluding current Science Cities

Whole sample Matched sample (1 nearest neighbour)

Outcome Raw difference T C ATT ATT b.a. Γ∗ (α= .05)

Patent applications, simple23.448***

71 6120.127*** 19.026***

2.75(8.978) (7.282) (6.982)

Patent applications, fractional7.254***

71 616.152*** 5.706***

2.55(2.703) (1.876) (1.773)

Night lights20.101***

71 614.995** 3.141

2.15(2.318) (2.202) (1.975)

Population in 201082.854***

71 6120.893 16.240

1.80(25.398) (15.016) (13.030)

Graduate share in 20100.103***

71 610.052*** 0.045***

4.65(0.009) (0.010) (0.010)

PhD share in 20100.003***

71 610.002*** 0.002***

2.00(0.000) (0.000) (0.000)

Number of SMEs (thousands, all)2.356***

45 370.180 0.008

2.05(0.900) (0.683) (0.606)

Number of SMEs (thousands, manufacturing)0.314**

45 370.118 0.103

1.65(0.125) (0.099) (0.094)

SMEs per 1000 people (all)-2.639*

45 37-1.173 -1.947

1.00(1.347) (1.660) (1.801)

SMEs per 1000 people (manufacturing)0.095

45 37-0.376 -0.302

1.00(0.128) (0.250) (0.257)

SME labour productivity (all)0.799***

45 370.318*** 0.329***

3.00(0.097) (94.352) (91.068)

SME labour productivity (manufacturing)0.636***

45 370.214 0.250*

2.30(0.100) (0.131) (0.128)

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is the p-value associated to each parameterestimate (standard errors are reported in parentheses). In the matched sample, T is the number of matched treated observa-tions; C is the number of matched controls; ‘ATT’ and ‘ATT b.a.’ are two estimates of (3) respectively excluding and includinga bias-adjustment term as per Abadie and Imbens (2011) (in both cases, standard errors are computed following Abadie andImbens (2006)); while Γ∗ is the minimum value of parameter Γ ≥ 1, selected from a grid spaced by intervals of 0.05 length,such that in a sensitivity analysis à la Rosenbaum (2002) the set of Wilcoxon signed-rank tests associated with Γ∗ do not si-multaneously reject the null hypothesis that the outcome variable is not different across the treated and control samples, fortests with α= 0.05 type I error. A higher value of Γ is associated to a stronger simulated unobserved factor which affects boththe outcome and the probability of receiving the treatment. Complete results of the sensitivity analysis for specific outcomesand for different values of Γ are available upon request.

48

Table 9: Firm-level Analysis: Probit estimation of model with latent variable representation (4),for different outcomes I f r

(1) (2) (3) (4)Technological Product Process

Does R&D innovation innovation innovationβ Coefficient -0.354 1.170∗∗ 0.912 0.520

(0.982) (0.586) (0.743) (0.804)

(Log) employees 0.296∗∗∗ 0.285∗∗∗ 0.221∗∗∗ 0.226∗∗∗(0.056) (0.042) (0.049) (0.045)

Young firms (0-5 years) -0.036 -0.321∗∗∗ -0.317∗∗∗ -0.330∗∗∗(0.117) (0.097) (0.112) (0.103)

50% or more foreign owned -0.167 -0.016 -0.311 -0.049(0.348) (0.350) (0.429) (0.434)

50% or more state owned -0.718 -0.961∗∗∗ -0.120 -1.559∗∗∗(0.533) (0.359) (0.411) (0.526)

Direct exporter 0.702∗∗∗ 0.231 0.429∗∗ 0.099(0.205) (0.156) (0.176) (0.160)

Main market: local 0.131 0.675∗ 0.336 1.354∗∗(0.348) (0.377) (0.420) (0.683)

Main market: national 1.038∗∗∗ 1.122∗∗∗ 0.772∗ 1.779∗∗∗(0.313) (0.391) (0.430) (0.676)

% of employees with a 0.012∗∗∗ 0.005∗∗∗ 0.006∗∗∗ 0.003∗completed university degree (0.002) (0.002) (0.002) (0.002)

Manufacturing 1.018∗∗∗ 1.424∗∗∗ 1.501∗∗∗ 1.299∗∗∗(0.135) (0.094) (0.118) (0.102)

Located in a city with -0.477∗∗ -0.148 -0.148 -0.003population over 1 million (0.211) (0.169) (0.188) (0.167)Region fixed effects YES YES YES YESObservations 4040 4040 4040 4040

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p < 0.01; where p is thep-value associated to each parameter estimate (standard errors are reported in parenthe-ses). All models include fixed effects for the Russian regions at the oblast administrativelevel. Standard errors are clustered at the municipal (rayon) level.

49

Table 10: Firm-level Analysis: OLS estimation of Equation (5) for different outcomes P f r

(1) (2) (3)(Log) sales Log(Y/L) Inn. share

of salesβ Coefficient -0.080 -0.491 -2.618

(0.132) (0.498) (3.821)

(Log) employees 0.173∗∗∗ 0.294(0.029) (0.252)

(Log) capital 0.126∗∗∗(0.017)

(Log) materials 0.752∗∗∗(0.031)

Young firms (0-5 years) 0.072∗∗ 0.089 -0.825∗(0.029) (0.060) (0.448)

50% or more foreign owned 0.055 0.551∗∗∗ 1.370(0.045) (0.183) (2.096)

50% or more state owned -0.185∗ -0.241 -1.082(0.098) (0.270) (2.955)

Direct exporter 0.043 0.627∗∗∗ 1.566∗(0.027) (0.079) (0.854)

Main market: local 0.094 0.130 2.425(0.058) (0.322) (2.296)

Main market: national 0.082 0.465 4.489∗(0.061) (0.324) (2.397)

% of employees with a 0.001∗∗∗ -0.001 0.024∗∗∗completed university degree (0.000) (0.001) (0.007)

Manufacturing -0.134∗∗∗ -0.167∗∗∗ 4.090∗∗∗(0.019) (0.046) (0.645)

Located in a city with 0.025 0.039 -1.212population over 1 million (0.046) (0.122) (1.152)Region fixed effects YES YES YESObservations 2809 2856 3933

Notes: ∗ denotes p < 0.10, ∗∗ denotes p < 0.05, and ∗∗∗ denotes p <0.01; where p is the p-value associated to each parameter estimate(standard errors are reported in parentheses). All models include fixedeffects for the Russian regions at the oblast administrative level. Stan-dard errors are clustered at the municipal (rayon) level. Differencesin sample size across models reflect differences in coverage about ac-counting data (employees, capital, materials).

50

ASc

ien

ceC

itie

sin

the

Sovi

etU

nio

nan

dR

uss

ia

Tab

leA

.1:S

cien

ceci

ties

Nau

kogr

adC

lose

dci

tyd

Pri

ori

tysp

ecia

lisa

tio

nar

easa

No.

Lo

cati

on

aO

bla

st

Foundede

YearSovietstatuse

YearRussianstatusf

Typea

Pas

tN

ow

Militarybase

Airrocketandspace

research

Electronicsandradio

engineering

Automation,ITand

instrumentation

Chemistry,chemicalphysics

andnewmaterials

Nuclearcomplex

Energetics

Biology,biotechnologyand

agriculturalsciences

1B

iysk

Alt

aiK

rai

1718

1957

2005

1N

oN

oN

oN

oN

oN

oYe

sN

oN

oYe

s

2M

irn

yA

rkh

ange

lsk

1957

1966

2Ye

sYe

sN

oYe

sN

oN

oN

oN

oN

oN

o

3Se

vero

dvi

nsk

Ark

han

gels

k19

3619

392

Yes

No

Yes

No

No

Yes

No

No

No

No

4Z

nam

ensk

Ast

rakh

an19

4819

622

Yes

Yes

Yes

Yes

No

No

No

No

No

No

5M

iass

Ch

elya

bin

sk17

7319

551

No

No

No

Yes

No

Yes

No

No

No

No

6O

zyo

rsk

Ch

elya

bin

sk19

4519

452

Yes

Yes

No

No

No

No

No

Yes

No

No

7Sn

ezh

insk

Ch

elya

bin

sk19

5719

572

Yes

Yes

No

No

No

No

No

Yes

No

No

8Tr

yokh

gorn

yC

hel

yab

insk

1952

1952

2Ye

sYe

sN

oN

oN

oYe

sN

oYe

sN

oN

o

9U

st-K

atav

Ch

elya

bin

sk17

5819

421

No

No

No

Yes

No

Yes

No

No

No

No

10K

asp

iysk

cD

ages

tan

1932

1936

2N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

11A

kad

emgo

rod

ok

Irku

tsk

1949

1988

5N

oN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sYe

s

12A

nga

rskc

Irku

tsk

1948

1957

2N

oN

oN

oN

oN

oN

oYe

sYe

sN

oN

o

13O

bn

insk

Kal

uga

1946

1946

2000

2N

oN

oN

oN

oN

oYe

sN

oYe

sYe

sN

o

14So

sen

skyc

Kal

uga

1952

1973

3N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

15K

om

som

ols

k-o

n-A

mu

rK

hab

arov

sk19

3219

341

No

No

No

Yes

No

Yes

No

No

No

No

16K

rasn

od

ar-5

9bK

rasn

od

arN

oN

oN

oN

oN

oN

oN

oN

oN

o

17A

kad

emgo

rod

ok

Kra

snoy

arsk

1944

1965

5N

oN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sYe

s

aB

ased

on

Agu

irre

chu

(200

9),u

nle

sssp

ecifi

edo

ther

wis

e.Ty

pe:

1-

Scie

nce

Cit

ies,

“sci

enti

fic

core

”es

tab

lish

edin

exis

tin

gci

ties

,wh

ich

oft

enh

ada

par

ticu

lar

his

tori

cals

ign

ifi-

can

ce;2

-Sci

ence

Cit

ies

that

rece

ived

city

stat

us

sim

ult

aneo

usl

y(o

ra

few

year

sla

ter)

wit

hth

ecr

eati

on

ofa

scie

nti

fic

or

scie

nti

fic-

ind

ust

rial

com

ple

xat

ap

ract

ical

lyn

ewlo

cati

on

(oft

enin

the

“op

enfi

eld

”);3

-Sc

ien

ceC

itie

sth

ath

ave

aris

enin

exis

tin

gse

ttle

men

tsan

dre

ceiv

edci

tyst

atu

saf

ter

ob

tain

ing

scie

nti

fic

fun

ctio

ns;

4-

Scie

nce

Cit

ies

that

do

no

t

hav

eci

tyst

atu

s;5

-ac

adem

icto

wn

.b

Bas

edo

nN

AS

(200

2).

cB

ased

on

Lap

po

and

Poly

an(2

008)

.d

Ru

ssia

nW

ikip

edia

arti

cle

on

clo

sed

citi

es(Z

ATO

),28

Sep

tem

ber

2016

.e

Wik

iped

iaar

ticl

esfo

rea

chci

ty,2

8Se

pte

mb

er20

16.

fR

uss

ian

Wik

iped

iaar

ticl

eo

nSc

ien

ceC

itie

s,28

Sep

tem

ber

2016

.

A.1

Tab

leA

.1–

con

tin

ued

fro

mp

revi

ou

sp

age

Nau

kogr

adC

lose

dci

tyd

Pri

ori

tysp

ecia

lisa

tio

nar

easa

No.

Lo

cati

on

aO

bla

st

Foundede

YearSovietstatuse

YearRussianstatusf

Typea

Pas

tN

ow

Militarybase

Airrocketandspace

research

Electronicsandradio

engineering

Automation,ITand

instrumentation

Chemistry,chemicalphysics

andnewmaterials

Nuclearcomplex

Energetics

Biology,biotechnologyand

agriculturalsciences

18Z

elen

ogo

rsk

Kra

snoy

arsk

1956

1956

2Ye

sYe

sN

oN

oN

oN

oYe

sYe

sN

oN

o

19Z

hel

ezn

ogo

rsk

Kra

snoy

arsk

1950

1954

2Ye

sYe

sN

oYe

sN

oN

oN

oYe

sN

oN

o

20K

urc

hat

ovc

Ku

rsk

1968

1976

2N

oN

oN

oN

oN

oN

oN

oYe

sYe

sN

o

21G

atch

ina

Len

ingr

ad19

2819

561

No

No

No

No

No

Yes

No

Yes

No

No

22P

rim

ors

kLe

nin

grad

1268

1948

1N

oN

oN

oYe

sN

oN

oN

oN

oN

oN

o

23So

snov

yB

or

Len

ingr

ad19

5819

623

Yes

Yes

No

No

Yes

No

No

Yes

Yes

No

24Z

elen

ogr

adM

osc

owC

ity

1958

1958

2N

oN

oN

oN

oYe

sYe

sN

oN

oN

oN

o

25A

vto

po

ligo

nM

osc

owO

bla

st19

6419

644

No

No

No

No

No

Yes

No

No

No

No

26B

alas

hik

ha

Mo

scow

Ob

last

1830

1942

1N

oN

oN

oYe

sN

oYe

sN

oN

oN

oN

o

27B

elo

ozer

sky

Mo

scow

Ob

last

1961

1961

4N

oN

oN

oYe

sN

oN

oN

oN

oN

oN

o

28C

her

no

golo

vka

Mo

scow

Ob

last

1710

1956

2008

3N

oN

oN

oN

oN

oYe

sYe

sN

oN

oN

o

29D

olg

op

rud

ny

Mo

scow

Ob

last

1931

1951

2N

oN

oN

oN

oN

oYe

sYe

sN

oN

oN

o

30D

ub

na

Mo

scow

Ob

last

1956

1956

2001

2N

oN

oN

oYe

sN

oYe

sN

oYe

sN

oN

o

31D

zerz

hin

sky

Mo

scow

Ob

last

1938

1956

3N

oN

oN

oYe

sN

oYe

sYe

sN

oN

oN

o

32Fr

yazi

no

Mo

scow

Ob

last

1584

1953

2003

3N

oN

oN

oYe

sYe

sYe

sN

oN

oN

oN

o

33Is

tra

Mo

scow

Ob

last

1589

1946

1N

oN

oN

oYe

sN

oYe

sN

oN

oYe

sN

o

34K

him

kiM

osc

owO

bla

st18

5019

501

No

No

No

Yes

No

Yes

No

No

No

No

35K

lim

ovsk

Mo

scow

Ob

last

1882

1940

1N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

36K

oro

lyov

Mo

scow

Ob

last

1938

1946

2001

1N

oN

oN

oYe

sN

oYe

sYe

sN

oN

oN

o

37K

rasn

oar

mey

skM

osc

owO

bla

st19

2819

341

No

No

No

No

No

Yes

Yes

No

No

No

aB

ased

on

Agu

irre

chu

(200

9),u

nle

sssp

ecifi

edo

ther

wis

e.Ty

pe:

1-

Scie

nce

Cit

ies,

“sci

enti

fic

core

”es

tab

lish

edin

exis

tin

gci

ties

,wh

ich

oft

enh

ada

par

ticu

lar

his

tori

cals

ign

ifi-

can

ce;2

-Sci

ence

Cit

ies

that

rece

ived

city

stat

us

sim

ult

aneo

usl

y(o

ra

few

year

sla

ter)

wit

hth

ecr

eati

on

ofa

scie

nti

fic

or

scie

nti

fic-

ind

ust

rial

com

ple

xat

ap

ract

ical

lyn

ewlo

cati

on

(oft

enin

the

“op

enfi

eld

”);3

-Sc

ien

ceC

itie

sth

ath

ave

aris

enin

exis

tin

gse

ttle

men

tsan

dre

ceiv

edci

tyst

atu

saf

ter

ob

tain

ing

scie

nti

fic

fun

ctio

ns;

4-

Scie

nce

Cit

ies

that

do

no

t

hav

eci

tyst

atu

s;5

-ac

adem

icto

wn

.b

Bas

edo

nN

AS

(200

2).

cB

ased

on

Lap

po

and

Poly

an(2

008)

.d

Ru

ssia

nW

ikip

edia

arti

cle

on

clo

sed

citi

es(Z

ATO

),28

Sep

tem

ber

2016

.e

Wik

iped

iaar

ticl

esfo

rea

chci

ty,2

8Se

pte

mb

er20

16.

fR

uss

ian

Wik

iped

iaar

ticl

eo

nSc

ien

ceC

itie

s,28

Sep

tem

ber

2016

.

A.2

Tab

leA

.1–

con

tin

ued

fro

mp

revi

ou

sp

age

Nau

kogr

adC

lose

dci

tyd

Pri

ori

tysp

ecia

lisa

tio

nar

easa

No.

Lo

cati

on

aO

bla

st

Foundede

YearSovietstatuse

YearRussianstatusf

Typea

Pas

tN

ow

Militarybase

Airrocketandspace

research

Electronicsandradio

engineering

Automation,ITand

instrumentation

Chemistry,chemicalphysics

andnewmaterials

Nuclearcomplex

Energetics

Biology,biotechnologyand

agriculturalsciences

38K

rasn

ogo

rskc

Mo

scow

Ob

last

1932

1942

3N

oN

oN

oN

oYe

sN

oN

oN

oN

oN

o

39K

rasn

ozn

amen

skM

osc

owO

bla

st19

5019

502

Yes

Yes

No

Yes

No

No

No

No

No

No

40Lu

khov

itsy

Mo

scow

Ob

last

1594

1957

?1

No

No

No

Yes

No

No

No

No

No

No

41Ly

tkar

ino

Mo

scow

Ob

last

1939

1957

1N

oN

oN

oYe

sN

oYe

sN

oN

oN

oN

o

42Ly

ub

erts

icM

osc

owO

bla

st16

2319

481

No

No

No

No

No

Yes

No

No

No

No

43M

end

eley

evo

Mo

scow

Ob

last

1957

1965

4N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

44M

ytis

hch

icM

osc

owO

bla

st14

6019

351

No

No

No

No

No

Yes

No

No

No

No

45O

bo

len

skM

osc

owO

bla

st19

7519

754

Yes

No

No

No

No

No

No

No

No

Yes

46O

revo

Mo

scow

Ob

last

1954

1954

4N

oN

oN

oYe

sN

oN

oN

oN

oN

oN

o

47Pe

resv

etM

osc

owO

bla

st19

4819

482

No

No

No

Yes

No

No

No

No

No

No

48P

rotv

ino

Mo

scow

Ob

last

1960

1960

2008

3Ye

sN

oN

oN

oN

oYe

sN

oYe

sN

oN

o

49P

ush

chin

oM

osc

owO

bla

st19

5619

6620

053

No

No

No

No

No

No

No

No

No

Yes

50R

emm

ash

Mo

scow

Ob

last

1957

1957

4N

oN

oN

oYe

sN

oN

oN

oN

oN

oN

o

51R

euto

vM

osc

owO

bla

st14

92-1

495

1940

2003

1N

oN

oN

oYe

sN

oYe

sN

oN

oN

oN

o

52To

mil

ino

Mo

scow

Ob

last

1894

1961

4N

oN

oN

oYe

sN

oYe

sN

oN

oN

oN

o

53Tr

oit

skM

osc

owO

bla

st16

1719

7720

073

No

No

No

No

No

Yes

No

Yes

No

No

54Yu

bile

yny

Mo

scow

Ob

last

1939

1950

3Ye

sN

oN

oYe

sN

oN

oN

oN

oN

oN

o

55Z

hel

ezn

od

oro

zhn

yM

osc

owO

bla

st18

6119

523

No

No

No

No

No

Yes

No

No

No

No

56Z

hu

kovs

kyM

osc

owO

bla

st19

3319

4720

072

No

No

No

Yes

No

No

No

No

No

No

57Z

vyoz

dn

ygo

rod

ok

Mo

scow

Ob

last

1960

1960

4Ye

sYe

sN

oYe

sN

oN

oN

oN

oN

oN

o

aB

ased

on

Agu

irre

chu

(200

9),u

nle

sssp

ecifi

edo

ther

wis

e.Ty

pe:

1-

Scie

nce

Cit

ies,

“sci

enti

fic

core

”es

tab

lish

edin

exis

tin

gci

ties

,wh

ich

oft

enh

ada

par

ticu

lar

his

tori

cals

ign

ifi-

can

ce;2

-Sci

ence

Cit

ies

that

rece

ived

city

stat

us

sim

ult

aneo

usl

y(o

ra

few

year

sla

ter)

wit

hth

ecr

eati

on

ofa

scie

nti

fic

or

scie

nti

fic-

ind

ust

rial

com

ple

xat

ap

ract

ical

lyn

ewlo

cati

on

(oft

enin

the

“op

enfi

eld

”);3

-Sc

ien

ceC

itie

sth

ath

ave

aris

enin

exis

tin

gse

ttle

men

tsan

dre

ceiv

edci

tyst

atu

saf

ter

ob

tain

ing

scie

nti

fic

fun

ctio

ns;

4-

Scie

nce

Cit

ies

that

do

no

t

hav

eci

tyst

atu

s;5

-ac

adem

icto

wn

.b

Bas

edo

nN

AS

(200

2).

cB

ased

on

Lap

po

and

Poly

an(2

008)

.d

Ru

ssia

nW

ikip

edia

arti

cle

on

clo

sed

citi

es(Z

ATO

),28

Sep

tem

ber

2016

.e

Wik

iped

iaar

ticl

esfo

rea

chci

ty,2

8Se

pte

mb

er20

16.

fR

uss

ian

Wik

iped

iaar

ticl

eo

nSc

ien

ceC

itie

s,28

Sep

tem

ber

2016

.

A.3

Tab

leA

.1–

con

tin

ued

fro

mp

revi

ou

sp

age

Nau

kogr

adC

lose

dci

tyd

Pri

ori

tysp

ecia

lisa

tio

nar

easa

No.

Lo

cati

on

aO

bla

st

Foundede

YearSovietstatuse

YearRussianstatusf

Typea

Pas

tN

ow

Militarybase

Airrocketandspace

research

Electronicsandradio

engineering

Automation,ITand

instrumentation

Chemistry,chemicalphysics

andnewmaterials

Nuclearcomplex

Energetics

Biology,biotechnologyand

agriculturalsciences

58A

pat

ity

(Aka

dem

goro

do

k)M

urm

ansk

1926

1954

2N

oN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sYe

s

59Po

lyar

nye

Zo

ric

Mu

rman

sk19

6819

732

No

No

No

No

No

No

No

Yes

Yes

No

60B

alak

hn

a(P

ravd

insk

)N

izh

ny

Nov

goro

d19

3219

411

No

No

No

No

Yes

No

No

No

No

No

61D

zerz

hin

skN

izh

ny

Nov

goro

d16

0619

302

No

No

No

No

No

No

Yes

No

No

No

62Sa

rov

Niz

hn

yN

ovgo

rod

1310

1947

1Ye

sYe

sN

oN

oN

oN

oN

oYe

sN

oN

o

63A

kad

emgo

rod

ok

Nov

osi

bir

sk19

5719

575

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

64K

olt

sovo

Nov

osi

bir

sk19

7919

7920

034

Yes

No

No

No

No

No

No

No

No

Yes

65K

rasn

oo

bsk

Nov

osi

bir

sk19

7019

784

No

No

No

No

No

No

No

No

No

Yes

66N

ovo

sib

irsk

-49b

Nov

osi

bir

skN

oN

oN

oN

oN

oN

oN

oN

oN

o

67O

msk

-5b

Om

skN

oN

o

68Z

arec

hn

yPe

nza

1954

1958

2Ye

sYe

sN

oN

oN

oN

oN

oYe

sYe

sN

o

69Pe

rm-6

bPe

rmN

oN

o

70B

ols

hoy

Kam

enc

Pri

mo

rsk

Kra

i19

4719

542

Yes

Yes

No

No

No

Yes

No

No

No

No

71Vo

lgo

do

nsk

cR

ost

ov19

5019

763

No

No

No

No

No

No

No

Yes

Yes

No

72Z

ern

ogr

adR

ost

ov19

2919

351

No

No

Yes

No

No

No

No

No

No

Yes

73Pe

terg

of

Sain

tPet

ersb

urg

1711

1960

2005

1N

oN

oN

oN

oN

oYe

sYe

sN

oN

oN

o

74D

esn

ogo

rskc

Smo

len

sk19

7419

822

No

No

No

No

No

No

No

Yes

Yes

No

75Le

snoy

Sver

dlo

vsk

1947

1954

2Ye

sYe

sN

oN

oN

oN

oN

oYe

sN

oN

o

76N

izh

nya

yaSa

lda

Sver

dlo

vsk

1760

1958

1N

oN

oN

oYe

sN

oN

oN

oN

oN

oN

o

77N

ovo

ura

lsk

Sver

dlo

vsk

1941

1949

2Ye

sYe

sN

oN

oN

oN

oYe

sYe

sN

oN

o

aB

ased

on

Agu

irre

chu

(200

9),u

nle

sssp

ecifi

edo

ther

wis

e.Ty

pe:

1-

Scie

nce

Cit

ies,

“sci

enti

fic

core

”es

tab

lish

edin

exis

tin

gci

ties

,wh

ich

oft

enh

ada

par

ticu

lar

his

tori

cals

ign

ifi-

can

ce;2

-Sci

ence

Cit

ies

that

rece

ived

city

stat

us

sim

ult

aneo

usl

y(o

ra

few

year

sla

ter)

wit

hth

ecr

eati

on

ofa

scie

nti

fic

or

scie

nti

fic-

ind

ust

rial

com

ple

xat

ap

ract

ical

lyn

ewlo

cati

on

(oft

enin

the

“op

enfi

eld

”);3

-Sc

ien

ceC

itie

sth

ath

ave

aris

enin

exis

tin

gse

ttle

men

tsan

dre

ceiv

edci

tyst

atu

saf

ter

ob

tain

ing

scie

nti

fic

fun

ctio

ns;

4-

Scie

nce

Cit

ies

that

do

no

t

hav

eci

tyst

atu

s;5

-ac

adem

icto

wn

.b

Bas

edo

nN

AS

(200

2).

cB

ased

on

Lap

po

and

Poly

an(2

008)

.d

Ru

ssia

nW

ikip

edia

arti

cle

on

clo

sed

citi

es(Z

ATO

),28

Sep

tem

ber

2016

.e

Wik

iped

iaar

ticl

esfo

rea

chci

ty,2

8Se

pte

mb

er20

16.

fR

uss

ian

Wik

iped

iaar

ticl

eo

nSc

ien

ceC

itie

s,28

Sep

tem

ber

2016

.

A.4

Tab

leA

.1–

con

tin

ued

fro

mp

revi

ou

sp

age

Nau

kogr

adC

lose

dci

tyd

Pri

ori

tysp

ecia

lisa

tio

nar

easa

No.

Lo

cati

on

aO

bla

st

Foundede

YearSovietstatuse

YearRussianstatusf

Typea

Pas

tN

ow

Militarybase

Airrocketandspace

research

Electronicsandradio

engineering

Automation,ITand

instrumentation

Chemistry,chemicalphysics

andnewmaterials

Nuclearcomplex

Energetics

Biology,biotechnologyand

agriculturalsciences

78Ve

rhn

yaya

Sald

acSv

erd

lovs

k17

7819

333

No

No

No

No

No

No

Yes

No

No

No

79Z

arec

hn

ySv

erd

lovs

k19

5519

552

No

No

No

No

No

No

No

Yes

Yes

No

80M

ich

uri

nsk

Tam

bov

1635

1932

2003

1N

oN

oN

oN

oN

oYe

sN

oN

oN

oYe

s

81Z

elen

od

ols

kcTa

tars

tan

1865

1949

1N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

82A

kad

emgo

rod

ok

Tom

sk19

7219

725

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

83Se

vers

kTo

msk

1949

1949

2Ye

sYe

sN

oN

oN

oN

oYe

sYe

sN

oN

o

84R

edki

no

Tve

r18

4319

504

No

No

No

No

Yes

No

Yes

No

No

No

85So

lnec

hn

yT

ver

1947

1951

4Ye

sYe

sYe

sN

oN

oYe

sN

oN

oN

oN

o

86U

do

mly

acT

ver

1478

1984

3N

oN

oN

oN

oN

oN

oN

oYe

sYe

sN

o

87G

lazo

vcU

dm

urt

ia16

7819

481

No

No

No

No

No

No

No

Yes

No

No

88Vo

tkin

skc

Ud

mu

rtia

1759

1957

1N

oN

oN

oYe

sN

oYe

sN

oN

oN

oN

o

89D

imit

rovg

rad

Uly

anov

sk16

9819

561

No

No

No

No

No

No

No

Yes

Yes

No

90K

ovro

vV

lad

imir

1778

1916

1N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

91M

elen

kiV

lad

imir

1778

1N

oN

oN

oN

oN

oYe

sN

oN

oN

oN

o

92R

adu

zhn

yV

lad

imir

1971

1971

2Ye

sYe

sN

oN

oYe

sYe

sN

oN

oN

oN

o

93N

ovov

oro

nez

hc

Voro

nez

h19

5719

642

No

No

No

No

No

No

No

Yes

Yes

No

94B

oro

kYa

rosl

avl

1807

1956

4N

oN

oN

oN

oN

oN

oN

oN

oN

oYe

s

95Pe

resl

avl-

Zal

essk

yYa

rosl

avl

1152

1964

1N

oN

oN

oN

oN

oYe

sYe

sN

oN

oN

o

aB

ased

on

Agu

irre

chu

(200

9),u

nle

sssp

ecifi

edo

ther

wis

e.Ty

pe:

1-

Scie

nce

Cit

ies,

“sci

enti

fic

core

”es

tab

lish

edin

exis

tin

gci

ties

,wh

ich

oft

enh

ada

par

ticu

lar

his

tori

cals

ign

ifi-

can

ce;2

-Sci

ence

Cit

ies

that

rece

ived

city

stat

us

sim

ult

aneo

usl

y(o

ra

few

year

sla

ter)

wit

hth

ecr

eati

on

ofa

scie

nti

fic

or

scie

nti

fic-

ind

ust

rial

com

ple

xat

ap

ract

ical

lyn

ewlo

cati

on

(oft

enin

the

“op

enfi

eld

”);3

-Sc

ien

ceC

itie

sth

ath

ave

aris

enin

exis

tin

gse

ttle

men

tsan

dre

ceiv

edci

tyst

atu

saf

ter

ob

tain

ing

scie

nti

fic

fun

ctio

ns;

4-

Scie

nce

Cit

ies

that

do

no

t

hav

eci

tyst

atu

s;5

-ac

adem

icto

wn

.b

Bas

edo

nN

AS

(200

2).

cB

ased

on

Lap

po

and

Poly

an(2

008)

.d

Ru

ssia

nW

ikip

edia

arti

cle

on

clo

sed

citi

es(Z

ATO

),28

Sep

tem

ber

2016

.e

Wik

iped

iaar

ticl

esfo

rea

chci

ty,2

8Se

pte

mb

er20

16.

fR

uss

ian

Wik

iped

iaar

ticl

eo

nSc

ien

ceC

itie

s,28

Sep

tem

ber

2016

.

A.5

BM

un

icip

alle

veld

ata

sou

rces

and

vari

able

s

Tab

leB

.1:M

un

icip

alle

veld

ata

sou

rces

and

vari

able

s

Dat

aty

pe

Dat

asu

b-t

ype

Dat

aso

urc

eD

escr

ipti

on

Fact

ors

guid

ing

the

sele

ctio

no

flo

cati

on

ofs

cien

ceci

ties

Ad

min

istr

ativ

eV

ario

us

iden

tifi

cati

on

info

rmat

ion

for

mu

nic

ipal

ity,

regi

on

and

fed

-

eral

dis

tric

t

Op

enSt

reet

Map

s,av

aila

ble

thro

ugh

GIS

-

LA

B(http://gis-lab.info/qa/osm-

adm.html

)

Un

iqu

em

un

icip

alit

y,fe

der

ald

istr

ict

and

regi

on

(ob

last

,kr

ai,

rep

ub

lic)

iden

tifi

cato

rs,

cod

esan

d

nam

es

Pop

ula

tio

n19

59ce

nsu

sd

ata

Jan

uar

y19

59So

viet

Cen

sus,

avai

l-

able

thro

ugh

Dem

osc

op

e(http:

//demoscope.ru/weekly/ssp/census.

php?cy=3

)

All

po

pu

lati

on

inm

un

icip

alit

yin

1959

,est

imat

esfo

r

som

em

un

icip

alit

ies

Geo

grap

hy

Are

aC

alcu

late

din

QG

ISb

ased

on

Op

enSt

reet

Map

s

Mu

nic

ipal

ity

area

calc

ula

ted

inQ

GIS

,m

easu

red

in

squ

ared

kilo

met

ers

Co

ord

inat

eso

fth

em

un

icip

alit

y

cen

tre

GP

Sco

ord

inat

eso

fth

ece

ntr

eo

fmu

nic

ipal

ity

calc

u-

late

din

QG

IS

Alt

itu

de

CG

IAR

.aC

on

sort

ium

for

Spat

ial

Info

rma-

tio

n(C

GIA

R-C

SI)

SRT

M90

mD

igit

alE

leva

-

tio

nD

ata,

vers

ion

4,av

aila

ble

athttp://

srtm.csi.cgiar.org/

Alt

itu

de

of

mu

nic

ipal

ity

inm

eter

s(m

ean

,m

edia

n,

SD,m

inan

dm

axva

lue)

Tem

per

atu

res

inJa

nu

ary

and

July

Wo

rld

Cli

mve

rsio

n1

(http://www.

worldclim.org/version1

),d

evel

op

ed

by

Hij

man

set

al.(

2005

)

Mo

nth

lyte

mp

erat

ure

dat

a,fo

rth

ep

erio

d19

60-

1990

,as

sign

edto

mu

nic

ipal

itie

sin

QG

IS.

Ave

rage

,

med

ian

,st

and

ard

dev

iati

on

,m

inim

um

,an

dm

axi-

mu

m.

Co

nti

nu

edo

nn

extp

age

A.6

Tab

leB

.1–

Co

nti

nu

edfr

om

pre

vio

us

pag

e

Dat

aty

pe

Dat

asu

b-t

ype

Dat

aso

urc

eD

escr

ipti

on

Rai

lro

adVe

rnad

sky

Stat

eG

eolo

gica

lM

use

um

and

U.S

.G

eolo

gica

lSu

rvey

,20

0106

00(2

001)

and

Cen

tral

man

agem

ent

un

ito

fth

em

il-

itar

yco

mm

un

icat

ion

so

fth

eR

edA

rmy

(194

3)

Dat

ao

nra

ilro

ads

wer

eco

nst

ruct

edu

sin

gra

ilro

ads

shap

efile

des

crib

ing

the

rail

road

so

fth

efo

rmer

So-

viet

Un

ion

aso

fth

eea

rly

1990

sp

rep

ared

by

Ver-

nad

sky

Stat

eG

eolo

gica

lM

use

um

and

U.S

.Geo

log-

ical

Surv

ey,

2001

0600

(200

1),

alo

ng

wit

ha

map

of

rail

road

sfr

om

1943

fro

mC

entr

alm

anag

emen

tu

nit

of

the

mil

itar

yco

mm

un

icat

ion

so

fth

eR

edA

rmy

(194

3)to

man

ual

lyre

mov

ean

yd

iffe

ren

ces

bet

wee

n

the

situ

atio

nd

epic

ted

inth

esh

apefi

lean

dth

e19

43

map

.In

dic

ato

req

ual

to1

ifm

un

icip

alit

yh

asac

cess

tora

ilro

ad,a

nd

0o

ther

wis

e.

Co

astl

ine/

maj

or

rive

r/la

keN

atu

ral

Ear

th,

1:10

mP

hys

ical

Vect

ors

ver-

sio

n(http://www.naturalearthdata.

com/downloads/10m-physical-

vectors/

)

Ind

icat

or

equ

alto

1if

mu

nic

ipal

ity

has

acce

ssto

coas

t/m

ajo

rri

ver/

lake

and

0o

ther

wis

e.

Dis

tan

ces

Cal

cula

ted

inQ

GIS

bas

edo

nth

eso

urc

es

spec

ified

abov

e

Dis

tan

ce(i

nkm

)fr

om

the

cen

tre

of

mu

nic

ipal

ity

to

the

nea

rest

railr

oad

,co

ast,

maj

or

rive

r,la

ke,

USS

R

bo

rder

,pla

nt(

ofa

ny

typ

e),a

nd

HE

I(o

fan

yty

pe)

.

Leve

lo

fin

du

stri

ald

evel

-

op

men

t

Dat

ao

nth

efa

cto

ries

,res

earc

han

d

des

ign

esta

blis

hm

ents

of

the

So-

viet

def

ence

ind

ust

ryin

1947

Dex

ter

and

Ro

dio

nov

(201

6).

Th

ed

atas

et

con

tain

sal

mo

st30

,000

entr

ies

and

in-

clu

des

the

nam

e,lo

cati

on

,mai

nb

ran

cho

f

def

ence

pro

du

ctio

n,e

stab

lish

men

tty

pe

as

wel

las

the

star

tan

den

dd

ate

for

the

esta

b-

lish

men

t’sm

ilita

ryw

ork

.

Nu

mb

ero

fall

pla

nts

,reg

ard

less

oft

ype

Hig

her

edu

cati

on

inst

i-

tute

s(H

EI)

HE

Isin

the

mu

nic

ipal

ity

in19

59D

eW

itt(

1961

)N

um

ber

of

all

HE

Is,

HE

Issp

ecia

lisn

gin

tech

nic

al

scie

nce

s,an

dH

EIs

spec

iali

sin

gin

bio

logy

and

med

-

ical

scie

nce

s

Co

nti

nu

edo

nn

extp

age

A.7

Tab

leB

.1–

Co

nti

nu

edfr

om

pre

vio

us

pag

e

Dat

aty

pe

Dat

asu

b-t

ype

Dat

aso

urc

eD

escr

ipti

on

R&

Din

stit

ute

sR

&D

inst

itu

tes

aso

f195

9V

ario

us

op

enso

urc

es(p

rim

arily

Wik

iped

ia)

Alo

wer

bo

un

dfo

rth

en

um

ber

of

R&

Din

stit

ute

s

(RA

S,n

on

-RA

San

dto

tal)

ho

sted

inm

un

icip

alit

yin

1959

-in

clu

des

on

lyth

ose

esta

bli

shed

bef

ore

or

in

1959

and

stil

lexi

stin

gn

ow.

RA

S-

Ru

ssia

nA

cad

emy

ofS

cien

ce.

Lo

ng-

term

ou

tco

mes

ofi

nte

rest

Pate

nts

Ap

plic

atio

ns

toE

PO

Eu

rop

ean

Pate

ntO

ffice

.Pat

ents

are

mat

ched

tom

un

icip

alit

ies

via

inve

nto

rs’

add

ress

es.

Nu

mb

ero

fp

aten

tsap

pli

cati

on

sto

EP

Oin

2006

-

2015

,by

inve

nto

r(s

imp

lean

dfr

acti

on

alco

un

tin

g)

Pop

ula

tio

n20

10ce

nsu

sd

ata

2010

Ru

ssia

nC

ensu

s,av

aila

ble

at

http://www.gks.ru/free_doc/new_

site/perepis2010/croc/perepis_

itogi1612.htm

All

po

pu

lati

on

,p

op

ula

tio

nw

ith

hig

her

edu

cati

on

,

and

po

pu

lati

on

wit

hP

hD

ord

oct

ora

ldeg

rees

inm

u-

nic

ipal

ity

in20

10

Nig

htt

ime

ligh

tsA

vera

gest

able

nig

htl

igh

tsVe

rsio

n4

DM

SP-O

LSN

igh

ttim

eLi

ghts

Tim

eSe

ries

,N

atio

nal

Oce

anic

and

Ato

msp

her

icA

dm

inis

trat

ion

(NO

AA

)

(http://ngdc.noaa.gov/eog/dmsp/

downloadV4composites.html

)

Ave

rage

nig

htl

igh

tsfo

r19

92-1

994

and

2009

-201

1,

clea

ned

ofg

asfl

ares

SME

sR

esu

lts

oft

he

2010

SME

cen

sus

Ro

ssta

t(F

eder

alSt

ate

Stat

isti

csSe

r-

vice

)(http://www.gks.ru/free_doc/

new_site/business/prom/small_

business/itog-spn.html

)

Th

ed

atas

etco

nta

ins

info

rmat

ion

on

the

nu

mb

er

of

firm

s,re

ven

ue,

nu

mb

ero

fem

plo

yees

,fi

xed

as-

sets

,an

dfi

xed

cap

ital

inve

stm

ent

by

size

,1-

and

4-

dig

itIS

ICse

cto

r.W

eu

seSM

Es

per

cap

ita,

SME

sale

s

per

wo

rker

(lab

ou

rp

rod

uct

ivit

y),a

nd

SME

sale

sp

er

un

ito

ffi

xed

lab

ou

r,al

lse

cto

rsan

dm

anu

fact

uri

ng

on

ly.

Th

eSM

Ece

nsu

sd

oes

no

tco

ver

ZAT

Os,

so16

scie

nce

citi

es,

wh

ich

are

also

ZAT

Os,

are

no

tco

v-

ered

.

aC

GIA

Ris

agl

ob

alp

artn

ersh

ipo

fres

earc

ho

rgan

isat

ion

sd

edic

ated

tore

du

cin

gp

over

tyan

dh

un

ger,

imp

rovi

ng

hu

man

hea

lth

and

nu

trit

ion

,an

den

han

cin

gec

on

-

syst

emre

sili

ence

thro

ugh

agri

cult

ura

lres

earc

h.

CG

IAR

-CSI

issp

atia

lsci

ence

com

mu

nit

yth

atfa

cili

tate

sC

GIA

R’s

inte

rnat

ion

alag

ricu

ltu

rald

evel

op

men

tre

sear

ch

usi

ng

spat

iala

nal

ysis

,GIS

,an

dre

mo

tese

nsi

ng:http://www.cgiar-csi.org/

.

A.8

C BEEPS V Russia

BEEPS is an enterprise survey whose objective is to gain an understanding of firms’ per-

ception of the environment in which they operate in order to be able to assess the con-

straints to private sector growth and enterprise performance. It covers topics related to

infrastructure, sales and supplies, degree of competition, land and permits, crime, fi-

nance, business-government relations, labour and establishment performance. BEEPS

is implemented by private contractors, using face-to-face interviews in the country’s of-

ficial language(s).

In BEEPS V, for the first time 37 Russian regions were covered, at least one in each

federal district. The survey was primarily targeted at top managers (CEOs), but in reality

the respondents often included accountants or operations managers. A total of 4,220

face-to-face interviews were completed, on average 114 interviews per region (see Ta-

ble C.1).

Table C.1: BEEPS V Russia sample breakdown

Region Number of interviews Region Number of interviewsCentral 1124 Siberian 709Belgorod 120 Irkutsk 131Kaluga 121 Kemerovo 124Kursk 87 Krasnoyarsk 89Lipetsk 121 Novosibirsk 123Moscow City 121 Omsk 120Moscow Oblast 122 Tomsk 122

Smolensk 71 Southern 328Tver 120 Krasnodar 88Voronezh 121 Rostov 120Yaroslavl 120 Volgograd 120Far Eastern 334 Urals 199Khabarovsk 122 Chelyabinsk 79Primorsky Krai 120 Sverdlovsk 120

Sakha (Yakutia) 92 Volga 922

North Caucasian 120 Bashkortostan 106Stavropol Krai 120 Kirov 134

Northwestern 484 Mordovia 120Kaliningrad 122 Nizhni Novgorod 82Leningrad 120 Perm 120Murmansk 120 Samara 120St. Petersburg 122 Tatarstan 120

Ulyanovsk 120Total 4220

Source: EBRD-World Bank BEEPS V Russia.

Also for the first time, BEEPS V Russia included an Innovation Module, with the aim

to obtain a better understanding of innovation - not only product innovation, but also

A.9

process, organisation and marketing innovation, as well as R&D and protection of in-

novation. The main questionnaire contained questions that determined eligibility for

participation in the Innovation Module, which was based on the third edition of the Oslo

Manual OECD and Statistical Office of the European Communities (2005). The so-called

filtering questions were asked with the help of show cards, which contained examples

of the relevant innovations to facilitate a common understanding of the definition of

innovation. While non-innovators did not receive additional questions on innovations,

innovating firms were asked to provide more information, including a detailed descrip-

tion of their main product or process innovation (in terms of impact on sales or costs

respectively).

Firms were only asked the relevant parts of the Innovation Module, which in turn col-

lected more detailed information on how the firms innovate, the level of innovativeness

and how important innovation is for the firms, as well as on R&D spending and patents.

Firms were asked to specify their main innovative product and process. More than 90

per cent of Innovation Module interviews were completed face-to-face immediately af-

ter the main questionnaire; 5.6 per cent were completed during a follow-up phone call,

and the rest during a second face-to-face visit or immediately after completing section

H in the main questionnaire. The interviews lasted on average 18.5 minutes.

The detailed descriptions of the firms’ main product or process innovation were used

to analyze whether the respective innovation complies with the formal definitions of

product and process innovation, taking into account the firm’s main business. Based on

this assessment, innovators may be reclassified as non-innovators, or moved to another

category of innovation than the one self-reported. As a result, about two thirds of the

self-reported innovations were reclassified, whereby 24 per cent were no longer classi-

fied as innovating firm, while the remaining innovations were reclassified according to

their type. The cleaning of innovations can only be done for product or process, i.e.

technological innovations, as no additional questions were asked for non-technological

innovations.

In Russia, only 51.9 per cent of companies that said they introduced new products

did product innovation and only 59.7 per cent of companies that said they introduced

new processes met the definition of process innovation. We also corrected the indicator

for R&D spending in the last three years based on the answers in the innovation mod-

ule. There was a significant variation across regions on all of these measures, which

could reflect both the competence of interviewers as well as understanding of the re-

A.10

spondents. We are not able to do the same for organizational and marketing innovation,

since we did not ask the respondents to name the main organisational and marketing

innovations.

D An Outline of the Soviet Innovation System

Throughout its history, the Soviet Union paid significant attention to R&D. Employment

in this sector grew from 35,000 in 1922 to more than 2.8 million in 1990, with intensive

investment made in R&D facilities and equipment along the way (Gokhberg, 1997). R&D

expenditures were also increasing in absolute terms until 1990, when they accounted for

about 3.5 per cent of GNP (Gokhberg and Mindely, 1993).

Despite these efforts, the specific characteristics of the Soviet economic and polit-

ical system harmed the efficiency of the R&D sector. In the absence of market forces,

R&D plans and findings were set as a result of bargaining among various parties in the

government and the R&D sector while the prices of R&D products were based on their

estimated costs. Following the branch-based structure of production and widespread

concerns for secrecy, the Soviet scientific sector was dominated by large multipurpose

R&D institutions able to exercise monopolistic power in the relevant fields (Schneider,

1994). General policies of autarky implemented by the Soviet Union were also applied to

science, resulting in the low level of technology and ideas exchange between the country

and the international community (Gokhberg, 1997).

These factors led to various inefficiencies in the sector. Deep governmental involve-

ment biased the concentration of Soviet R&D towards politically important areas such as

defence and engineering at the expense of fundamental research such as medical and

life sciences (Gokhberg and Mindely, 1993). Being subject to very specific incentives,

R&D efforts were often focused on obsolete areas and produced low-quality results. In

the struggle for funding, R&D institutions were reluctant to share their technologies and

were artificially increasing their staff by keeping older cohorts of researchers in employ-

ment.31 As a result of this disproportional increase in R&D personnel, increases in fund-

ing were mainly allocated to wages rather than to buying new state-of-the-art equip-

ment (Schneider, 1994). On the top of these problems, low incentives to innovate in the

production sector and weak diffusion of innovations across enterprises and industries

31For example, in the late 1980s almost 80 per cent of Soviet researchers were aged 50 years or more,compared to 34 per cent in the United States (Schneider, 1994).

A.11

hampered the utilization of the suboptimal output of the Soviet R&D.

R&D was largely separated from higher education, with universities becoming al-

most exclusively training centers. Basic research was instead concentrated in the sys-

tem of Academy of Sciences and branch academies of agricultural sciences, medical sci-

ences, and education. Applied R&D, on the other hand, was concentrated in the indus-

trial R&D units, which were established by each branch ministry. Enterprise R&D was

the least developed, as enterprises had no incentives to introduce new products and

processes. Hence, even in cases where the Soviet Union had a leading position in the

development of significant innovations, it fell behind others in diffusion of innovations.

These peculiar characteristics of Soviet R&D have to be taken into account while

analysing the development of the sector during the transition period. Quoting Gokhberg

(1997), “[o]nly a part of the R&D sector inherited from the Soviet era can and should be

preserved” (p. 9). This view was arguably shared by market reformers in the 1990’s, when

the Russian R&D sector went through a painful adjustment as part of Russia’s transfor-

mation into a market economy.

A.12

E Matching: Graphs and Additional Statistics

Figure E.1: Matching for non-SME oucomes; distribution of the p-score (scatter plots). Topgraph: distribution before matching. Bottom graph: distribution after matching.

A.13

Figure E.2: Matching for SME oucomes; distribution of the p-score (scatter plots). Top graph:distribution before matching. Bottom graph: distribution after matching.

A.14

Figure E.3: Matching for non-SME oucomes; distribution of the p-score (kernel densities). Topgraph: estimates before matching. Bottom graph: estimates after matching.

A.15

Figure E.4: Matching for SME oucomes; distribution of the p-score (kernel densities). Top graph:estimates before matching. Bottom graph: estimates after matching.

A.16

Table E.1: Covariate Balance: Mahalanobis Matching, all Science Cities

Matched sample: non-SME outcomesVariable Treated Control Std. bias t-stat. p-value V-ratioLatitude 55.664 55.778 -3.0 -0.20 0.843 0.90Longitude 49.771 49.506 1.2 0.08 0.936 1.13Average temperature in January (°C) -11.632 -11.875 6.6 0.43 0.667 1.21Average temperature in July (°C) 18.535 18.427 6.4 0.42 0.673 1.08Average altitude 0.1686 0.1829 -13.6 -0.90 0.372 0.71Minimum distance from coast 0.7246 0.7565 -7.7 -0.51 0.612 1.04Minimum distance from USSR border 0.6651 0.7034 -10.9 -0.72 0.475 0.92(Log) Population in 1959 2.8761 3.2188 -16.6 -1.09 0.276 1.11(Log) Number of universities in 1959 0.1910 0.1874 0.7 0.05 0.964 1.06(Log) Number of plants in 1947 1.1438 1.1514 -0.6 -0.04 0.967 1.00(Log) Number of R&D institutes in 1959 0.2899 0.2369 8.9 0.59 0.558 1.17(Log) Area in km2 -1.6808 -1.3622 -15.8 -1.04 0.301 0.84

Matched Sample: SME OutcomesVariable Treated Control Std. Bias t-stat. p-value V-ratioLatitude 55.664 56.087 -11.6 -0.70 0.483 1.05Longitude 49.771 46.498 16.0 0.95 0.342 1.49Average temperature in January (°C) -11.632 -11.556 -2.2 -0.13 0.897 1.49Average temperature in July (°C) 18.535 18.365 10.2 0.62 0.539 1.09Average altitude 0.1686 0.1689 -0.3 -0.02 0.987 0.94Minimum distance from coast 0.7246 0.7357 -2.8 -0.17 0.865 1.32Minimum distance from USSR border 0.6651 0.7057 -11.7 -0.71 0.479 1.01(Log) Population in 1959 2.8761 3.6303 -39.0 -2.32 0.022 1.63(Log) Number of universities in 1959 0.1910 0.2123 -4.0 -0.24 0.810 0.91(Log) Number of plants in 1947 1.1438 1.3668 -17.5 -1.07 0.287 0.85(Log) Number of R&D institutes in 1959 0.2899 0.3261 -5.7 -0.35 0.730 0.88(Log) Area in km2 -1.6808 -1.1679 -26.6 -1.62 0.108 0.94

Notes: For each variable, the table reports its mean for both the treated and control groups of the matchedsample, along with the corresponding standardised bias (the treated-control difference as a percentage ofthe combined sample’s standard deviation). In addition, for each variable the table reports the t-statisticand the associated p-value for the difference in the two means, as well as the ratio between the variancesof the treated and the variance of the control group for the variable in question. Both matched samplesare obtained via 1:1 Mahalanobis matching (with replacement), forcing exact matching on the followingdummy variables: presence of a major river, presence of a lake and presence of “historical” railroads in amunicipality.

A.17

Table E.2: Covariate Balance: Propensity Score Matching, Science Cities with no “official” statusin today’s Russia

Matched sample: non-SME outcomesVariable Treated Control Std. bias t-stat. p-value V-ratioLatitude 55.703 55.582 2.8 0.17 0.869 0.72Longitude 50.879 48.097 12.3 0.73 0.467 0.99Average temperature in January (°C) -11.888 -11.507 -7.3 -0.44 0.661 0.41Average temperature in July (°C) 18.516 18.609 -4.6 -0.28 0.784 0.69Average altitude 0.1725 0.1693 2.4 0.14 0.888 0.40Minimum distance from railroad 0.0071 0.0361 -29.1 -1.77 0.079 0.02Minimum distance from river 0.0336 0.0336 0.0 0.00 0.998 1.33Minimum distance from lake 0.1148 0.1249 -9.8 -0.59 0.559 0.51Minimum distance from coast 0.7237 0.6257 23.7 1.41 0.161 0.94Minimum distance from USSR border 0.6765 0.6451 8.3 0.49 0.622 0.90Population in 1959 74.326 64.129 9.2 0.55 0.586 1.92Number of universities in 1959 0.6216 0.3823 12.8 0.76 0.451 2.74Number of plants in 1947 6.9459 5.2206 13.4 0.79 0.430 1.82Number of R&D institutes in 1959 0.9054 0.4707 18.7 1.11 0.270 1.31Area in km2 0.8135 12.890 -18.6 -1.13 0.259 0.02

Matched Sample: SME OutcomesVariable Treated Control Std. Bias t-stat. p-value V-ratioLatitude 55.703 55.708 -0.1 -0.01 0.994 1.12Longitude 50.879 51.148 -1.1 -0.06 0.951 1.20Average temperature in January (°C) -11.888 -12.233 7.5 0.43 0.671 1.15Average temperature in July (°C) 18.516 18.515 0.1 0.00 0.998 1.14Average altitude 0.1725 0.1761 -2.9 -0.16 0.871 0.50Minimum distance from railroad 0.0071 0.0566 -24.7 -1.50 0.135 0.24Minimum distance from river 0.0336 0.0397 -17.0 -0.93 0.354 1.14Minimum distance from lake 0.1148 0.1188 -4.7 -0.26 0.796 0.95Minimum distance from coast 0.7237 0.7136 2.5 0.14 0.892 1.21Minimum distance from USSR border 0.6765 0.6892 -3.5 -0.19 0.846 1.07Population in 1959 74.326 76.104 -1.6 -0.09 0.931 1.95Number of universities in 1959 0.6216 0.6346 -0.6 -0.03 0.973 1.58Number of plants in 1947 6.9459 4.6538 19.5 1.03 0.306 3.42Number of R&D institutes in 1959 0.9054 0.3846 26.4 1.38 0.171 4.35Area in km2 0.8135 4.1237 -27.1 -1.65 0.102 0.25

Notes: For each variable, the table reports its mean for both the treated and control groups of thematched sample in question, along with the corresponding standardised bias (the treated-control dif-ference as a percentage of the combined standard deviation of the matched sample). In addition, foreach variable the table reports the t-statistic and the associated p-value about the difference in the twomeans, as well as the ratio between the variances of the treated and the variance of the control groupfor the variable in question. Both matched samples are obtained via 1:1 nearest neighbour matching onthe estimated propensity score (with replacement).

A.18

Table E.3: Covariate Balance: Mahalanobis Matching, Science Cities with no “official” status intoday’s Russia

Matched sample: non-SME outcomesVariable Treated Control Std. bias t-stat. p-value V-ratioLatitude 55.698 54.649 24.2 1.43 0.155 0.87Longitude 50.548 50.836 -1.3 -0.07 0.941 1.06Average temperature in January (°C) -11.829 -11.875 1.1 0.06 0.949 0.74Average temperature in July (°C) 18.520 18.546 -1.3 -0.07 0.941 0.67Average altitude 0.1714 0.2507 -32.1 -1.96 0.052 0.23Minimum distance from coast 0.7227 0.7108 2.8 0.16 0.870 0.95Minimum distance from USSR border 0.6767 0.6786 -0.5 -0.03 0.977 0.87(Log) Population in 1959 3.0551 3.6003 -31.6 -1.80 0.075 2.26(Log) Number of universities in 1959 0.2067 0.1183 19.2 1.10 0.273 2.78(Log) Number of plants in 1947 1.2676 0.8206 37.6 2.18 0.031 1.52(Log) Number of R&D institutes in 1959 0.3228 0.1169 36.6 2.09 0.038 2.38(Log) Area in km2 -1.3477 -0.2071 -63.0 -3.68 0.000 0.98

Matched sample: SME outcomesVariable Treated Control Std. bias T-stat. p-value V-ratioLatitude 55.703 55.828 -3.4 -0.17 0.862 1.26Longitude 50.879 46.070 24.2 1.22 0.227 1.73Average temperature in January (°C) -11.888 -11.543 -8.4 -0.44 0.659 0.86Average temperature in July (°C) 18.516 18.410 5.8 0.31 0.761 1.00Average altitude 0.1725 0.1912 -10.6 -0.61 0.541 0.22Minimum distance from coast 0.7237 0.7466 -5.7 -0.30 0.768 1.25Minimum distance from USSR border 0.6765 0.7109 -9.7 -0.50 0.615 1.11(Log) Population in 1959 3.0083 3.6190 -35.0 -1.73 0.086 2.79(Log) Number of universities in 1959 0.2029 0.0996 23.0 1.13 0.260 3.25(Log) Number of plants in 1947 1.2224 0.8806 29.6 1.52 0.131 1.41(Log) Number of R&D institutes in 1959 0.3073 0.0630 48.0 2.33 0.022 6.00(Log) Area in km2 -1.3528 -0.42227 -50.2 -2.64 0.010 1.17

Notes: For each variable, the table reports its mean for both the treated and control groups of the matchedsample, along with the corresponding standardised bias (the treated-control difference as a percentage ofthe combined sample’s standard deviation). In addition, for each variable the table reports the t-statisticand the associated p-value about the difference in the two means, as well as the ratio between the vari-ances of the treated and the variance of the control group for the variable in question. Both matched sam-ples are obtained via 1:1 Mahalanobis matching (with replacement), forcing exact matching on the follow-ing dummy variables: presence of a major river, presence of a lake and presence of “historical” railroads ina municipality.

A.19