economic analysis of scientific publications and …...supplementary figure6 – probability of...

SUPPLEMENTARY INFORMATIONARTICLE NUMBER: 16020 | DOI: 10.1038/NENERGY.2016.20

NATURE ENERGY | www.nature.com/natureenergy

Supplementary Information

Advice for Energy R&D Policy from an Economic Analysis of Scientific Publications

February 2016

David PoppProfessorDepartment of Public Administration and International AffairsCenter for Policy ResearchThe Maxwell SchoolSyracuse University426 Eggers HallSyracuse, NY 13244-1020

and

Research AssociateNational Bureau of Economic Research

ph: 315-443-2482fax: 315-443-1075e-mail: [email protected] page: http://faculty.maxwell.syr.edu/dcpopp/index.html

Economic analysis of scientific publications and implications for energy research and development

http://dx.doi.org/10.1038/nenergy.2016.20


SUPPLEMENTARY INFORMATION DOI: 10.1038/NENERGY.2016.20

This document provides describes the data used, elaborates on methodology, and

provides additional results and robustness checks. The sections are as follows:

Supplementary Figures p. 1 Supplementary Tables p. 10Supplementary Notes:1. Web of Science Keyword Searches p. 302. Data Description p. 303. Identifying the Lag Length p. 334. Discussion of Instrument Validity p. 365. Detailed Results and Robustness Checks p. 386. Detailed Results and Methodology for Patent Citations p. 447. Additional Results on Diminishing Returns to R&D p. 46



SUPPLEMENTARY INFORMATIONDOI: 10.1038/NENERGY.2016.20

SUPPLEMENTARY FIGURES

Supplementary Figure 1 – Trends in Energy R&D and Publications by Country: Biofuels




2

Supplementary Figure 2 – Trends in Energy R&D and Publications by Country: Energy Efficiency




3

Supplementary Figure 3 – Trends in Energy R&D and Publications by Country: Solar Energy




4

Supplementary Figure 4 – Trends in Energy R&D and Publications by Country: Wind Energy




5

Supplementary Figure 5 – Cumulative effect of energy R&D on publications

Figure shows the cumulative effect of energy R&D on publications through year t+x, where xrepresents years since funding occurred, shown on the x-axis. Calculated using the first differenced model with instrumental variables for contemporary energy R&D




6

Supplementary Figure 6 – Probability of patent citation over time

A. Annual probability of patent citation

B. Cumulative probability of patent citation

Panel A shows the annual probability, and panel B the cumulative probability, of an article receiving a citation t years after publication.




7

Supplementary Figure 7 – Marginal effect of R&D per year: Models including R&D squared

Figure shows the marginal effect of energy R&D on publications in year t+x, where x represents years since funding occurred, shown on the x-axis. Calculated at different percentiles of energy R&D spending, using the first differenced model including R&D2 with instrumental variables for contemporary energy R&D.




8

Supplementary Figure 8 – Cumulative effect of R&D: Models including R&D squared

Figure shows the cumulative marginal effect of energy R&D on publications through year t+x,where x represents years since funding occurred, shown on the x-axis. Calculated at different percentiles of energy R&D spending, using the first differenced model including R&D2 with instrumental variables for contemporary energy R&D.




9

Supplementary Figure 9—Number of countries doubling energy R&D per year

Figure shows the number of countries whose energy R&D budget for a given technology is at least double that of the previous year.




10

SUPPLEMENTARY TABLES

Supplementary Table 1 – Keyword SearchesBiofuels#11 TS = ("biomass" NEAR/5 "electricit*" OR "biomass fuel*" OR

"biomass heat*" OR "biomass energy" OR "Bio feedstock*" OR "biofeedstock*" OR "Hydrotreated vegetable oil*" or "lignocellulosic biomass*" OR "cellulosic ethanol*" or "biomass to liquid*" OR "bio synthetic gas*" OR "algae-based fuel*" OR "landfill gas*" or "Biohydrogen production*" or "Biological hydrogen production*" or "bio energy" or "bioenergy" or "biofuel*" or "bio fuel*" or "biodiesel*" or "bio diesel*" or "biogas*" or "bio gas*" OR "Bio syngas*" or "bio oil" or "bio ethanol*" or "bioethanol*" OR "fuel ethanol*" OR "Biomethanol*" OR "bio methanol*") NOT TS = ("co-combust*" or "cocombust*" or "co-fir*" or "cofir*" or "multi-combust*" or "multicombust*" or "multi-fir*" or "multifir*" or "fuel cell*" or"biofuel cell*") Refined by: Web of Science Categories=( ENERGY FUELS OR SPECTROSCOPY OR BIOTECHNOLOGY APPLIED MICROBIOLOGY OR ENTOMOLOGY OR ENGINEERING CHEMICAL OR ENVIRONMENTAL SCIENCES OR POLYMER SCIENCE OR AGRICULTURAL ENGINEERING OR ENGINEERING ENVIRONMENTAL OR GEOSCIENCES MULTIDISCIPLINARY OR CHEMISTRY MULTIDISCIPLINARY OR TRANSPORTATION SCIENCE TECHNOLOGY OR FOOD SCIENCE TECHNOLOGY OR CHEMISTRY PHYSICAL OR CHEMISTRY APPLIED OR GENETICS HEREDITY OR BIOCHEMISTRY MOLECULAR BIOLOGY OR BIOLOGY OR WATER RESOURCES OR THERMODYNAMICS OR CHEMISTRY ORGANIC OR AGRONOMY OR PHYSICS ATOMIC MOLECULAR CHEMICAL OR GEOCHEMISTRY GEOPHYSICS OR PLANT SCIENCES OR ENGINEERING MECHANICAL OR CHEMISTRY ANALYTICAL OR MULTIDISCIPLINARY SCIENCES OR METEOROLOGY ATMOSPHERIC SCIENCES OR MATERIALS SCIENCE BIOMATERIALS OR AGRICULTURE MULTIDISCIPLINARY OR DEVELOPMENTAL BIOLOGY OR MICROBIOLOGY OR ECOLOGY OR MECHANICS OR ENGINEERING INDUSTRIAL OR FORESTRY OR HORTICULTURE OR BIOCHEMICAL RESEARCH METHODS OR NANOSCIENCE NANOTECHNOLOGY OR ENGINEERING MULTIDISCIPLINARY OR SOIL SCIENCE OR MATERIALS SCIENCE PAPER WOOD OR METALLURGY METALLURGICAL ENGINEERING OR MATERIALS SCIENCE TEXTILES OR ELECTROCHEMISTRY OR ENGINEERING CIVIL OR MATERIALS SCIENCE MULTIDISCIPLINARY ) Timespan=1991-2010. Databases=SCI-EXPANDED. Lemmatization=Off




11

Energy Efficiency

Note: The energy efficiency search strategy uses the three separate searches below, and excludes publications included in other energy categories,

#1 TS = ("waste heat" NEAR/3 (recover* OR convert* OR utilize* OR use)) OR TS = ("district heat*" OR "district cool*") OR TS = (("LED" OR "light emitting diode") NEAR/1 (lighting OR lightbulb* OR "light bulb*" OR lamp* OR “solid state light*” OR “solid state lamp*”)) OR TS = (("CFL" OR "compact fluorescent") NEAR/1 (lighting OR lightbulb* OR "light bulb*" OR lamp*)) OR TS = "solid state light*"

#2 (TS = ((electric OR hybrid OR “electric drive” OR “hybrid drive”) NEAR/0 (vehicle* OR automobile* OR "auto" OR "autos" OR"car" OR "cars")) NOT TS = (“fuel cell*” OR “fuel-cell*” OR “Hybrid Car-Parrinello”)) Refined by: [excluding] Web of Science Categories=( ASTRONOMY ASTROPHYSICS OR DERMATOLOGY OR INFECTIOUS DISEASES OR LITERATURE BRITISH ISLES OR MEDICINE GENERAL INTERNAL OR PHARMACOLOGY PHARMACY OR HISTORY PHILOSOPHY OF SCIENCE OR SPECTROSCOPY OR VETERINARY SCIENCES OR CLINICAL NEUROLOGY ) Timespan=1991-2010. Databases=SCI-EXPANDED. Lemmatization=Off




12

#3 (TS=((energy OR fuel OR gas* OR electric* OR petrol*) NEAR/1(consum* OR use OR using OR usage OR burn*) NEAR/1 (reduc* OR less OR lower OR decreas*)) OR TS=((energy OR fuel OR gas* OR petrol) NEAR/1 (efficien* OR economy OR mileage OR productivity) NEAR/1 (improv* OR increas* OR better OR greater OR more)) OR TS=((energy OR fuel OR gas* OR electric* OR petrol*) NEAR/1 (saving* OR save OR saves OR saved))) NOT TS = ("fuel cell*" OR “fuel-cell*”)Refined by: [excluding] Web of Science Categories=( ENVIRONMENTAL STUDIES OR FOOD SCIENCE TECHNOLOGY OR ZOOLOGY OR NUTRITION DIETETICS OR PHYSIOLOGY OR BIOCHEMISTRY MOLECULAR BIOLOGY OR BIOLOGY OR METEOROLOGY ATMOSPHERIC SCIENCES OR CARDIAC CARDIOVASCULAR SYSTEMS OR ASTRONOMY ASTROPHYSICS OR MATHEMATICS APPLIED OR ROBOTICS ) AND Web of Science Categories=( ENERGY FUELS OR ENGINEERING CHEMICAL OR ENGINEERING ELECTRICAL ELECTRONIC OR ENGINEERING MECHANICAL OR THERMODYNAMICS OR ENVIRONMENTAL SCIENCES OR CONSTRUCTION BUILDING TECHNOLOGY OR MATERIALS SCIENCE MULTIDISCIPLINARY OR ENGINEERING ENVIRONMENTAL OR METALLURGY METALLURGICAL ENGINEERING OR TELECOMMUNICATIONS OR COMPUTER SCIENCE INFORMATION SYSTEMS OR COMPUTER SCIENCE HARDWARE ARCHITECTURE OR COMPUTER SCIENCE THEORY METHODS OR TRANSPORTATION SCIENCE TECHNOLOGY OR ENGINEERING MULTIDISCIPLINARY OR MATERIALS SCIENCE PAPER WOOD OR AGRICULTURAL ENGINEERING OR AUTOMATION CONTROL SYSTEMS OR MATERIALS SCIENCE CERAMICS OR COMPUTER SCIENCE SOFTWARE ENGINEERING OR MINING MINERAL PROCESSING OR ENGINEERING INDUSTRIAL OR COMPUTER SCIENCE ARTIFICIAL INTELLIGENCE OR AGRONOMY OR MATERIALS SCIENCE COMPOSITES OR MATERIALS SCIENCE TEXTILES OR OPERATIONS RESEARCH MANAGEMENT SCIENCE OR ENGINEERING MARINE OR ENGINEERING OCEAN OR URBAN STUDIES )Databases=SCI-EXPANDED Timespan=1991-2010Lemmatization=Off




13

Solar Energy

The solar energy search combines searches for specific types of solar energy (e.g. solar PV) and a more general search strategy:

Solar Thermal Power#1 (TS = (solar NEAR/2 thermoelectr*) OR TS = (solar NEAR/2 “power

plant”) OR TS = (“concentrat* solar” NEAR/2 power) OR TS= (“solar thermal” NEAR/2 (power OR electric*)) OR TS=(parabolic* NEAR/2 trough*) OR TS=((parabolic NEAR/2 dish*) AND solar) OR TS = (stirling NEAR/2 dish*) OR TS=((Fresnel NEAR/2 (reflector* OR lens*)) AND solar)) NOT (TS = (cell* OR photovoltaic* OR PV) OR TS = (hydrogen NEAR/1 (generat* or product*)) OR TS = (battery OR batteries) OR TS = (storage OR store OR storing)) Refined by: [excluding] Web of Science Categories=( ENGINEERING AEROSPACE OR ASTRONOMY ASTROPHYSICS ) Timespan=1991-2010. Databases=SCI-EXPANDED. Lemmatization=Off

#2 TS=(solar NEAR/2 tower) NOT (TS = (cell* OR photovoltaic* OR PV) OR TS = (hydrogen NEAR/1 (generat* or product*)) OR TS = (battery OR batteries) OR TS = (storage OR store OR storing)) Refined by: [excluding] Web of Science Categories=( ASTRONOMY ASTROPHYSICS OR NUCLEAR SCIENCE TECHNOLOGY OR METEOROLOGY ATMOSPHERIC SCIENCES ) Timespan=1991-2010. Databases=SCI-EXPANDED. Lemmatization=Off

Solar Photovoltaic#3 TS = ("photovoltaic energ*" OR "solar cell*" OR "photovoltaic power

*" OR "photovoltaic cell*" OR "photovoltaic solar energy*") NOT (TS = (hydrogen NEAR/1 (generat* or product*)) OR TS = (battery OR batteries) OR TS = (storage OR store OR storing))




14

Solar General#4 TS = (“solar panel*” OR “solar array*” OR “solar resource*” OR

“solar potential” OR “solar energy” OR “solar collector*”) NOT (#5 OR #8 OR #9 OR TS = (hydrogen NEAR/1 (generat* or product*)) OR TS = (battery OR batteries) OR TS = (storage OR store OR storing)) Refined by: Web of Science Categories=( AUTOMATION CONTROL SYSTEMS OR CHEMISTRY ANALYTICAL OR CHEMISTRY INORGANIC NUCLEAR OR CHEMISTRY MULTIDISCIPLINARY OR CHEMISTRY ORGANIC OR CHEMISTRY PHYSICAL OR CONSTRUCTION BUILDING TECHNOLOGY OR ELECTROCHEMISTRY OR ENERGY FUELS OR ENGINEERING CIVIL OR ENGINEERING ELECTRICAL ELECTRONIC OR ENGINEERING MULTIDISCIPLINARY OR ENVIRONMENTAL SCIENCES OR IMAGING SCIENCE PHOTOGRAPHIC TECHNOLOGY OR MATERIALS SCIENCE CERAMICS OR MATERIALS SCIENCE COATINGS FILMS OR MATERIALS SCIENCE MULTIDISCIPLINARY OR MECHANICS OR METALLURGY METALLURGICAL ENGINEERING OR MINING MINERAL PROCESSING OR NANOSCIENCE NANOTECHNOLOGY OR OPTICS OR PHYSICS APPLIED OR PHYSICS CONDENSED MATTER OR PHYSICS NUCLEAR OR POLYMER SCIENCE OR THERMODYNAMICS OR WATER RESOURCES ) AND [excluding] Web of Science Categories=( METEOROLOGY ATMOSPHERIC SCIENCES OR ENGINEERING AEROSPACE OR ASTRONOMY ASTROPHYSICS)Timespan=1991-2010. Databases=SCI-EXPANDED. Lemmatization=Off




15

Wind Energy

#1 TS = ("wind power" OR "wind energy" OR "wind turbine*" OR "wind farm*" OR "wind park*" OR "wind plant*") NOT TS = (battery OR batteries OR storage OR store OR storing OR "hydrogen production*" OR "wind" NEAR "hydrogen" OR "grid integration*" OR "load management" OR "offshore" NEAR/5 ("connect*" OR "link*" OR "electric*" OR "grid*")) Refined by: Web of Science Categories=( ENERGY FUELS OR MATERIALS SCIENCE COMPOSITES OR ENGINEERING ELECTRICAL ELECTRONIC OR ORNITHOLOGY OR ENGINEERING MECHANICAL OR ENVIRONMENTAL SCIENCES OR COMPUTER SCIENCE ARTIFICIAL INTELLIGENCE OR MECHANICS OR MATERIALS SCIENCE CHARACTERIZATION TESTING OR ENGINEERING CIVIL OR PHYSICS MULTIDISCIPLINARY OR THERMODYNAMICS OR STATISTICS PROBABILITY OR MATHEMATICS INTERDISCIPLINARY APPLICATIONS OR METEOROLOGY ATMOSPHERIC SCIENCES OR ENGINEERING MARINE OR ENGINEERING MULTIDISCIPLINARY OR ECOLOGY OR METALLURGY METALLURGICAL ENGINEERING OR AUTOMATION CONTROL SYSTEMS OR INSTRUMENTS INSTRUMENTATION OR MATERIALS SCIENCE MULTIDISCIPLINARY OR MULTIDISCIPLINARY SCIENCES OR BIOLOGY OR PHYSICS APPLIED OR COMPUTER SCIENCE THEORY METHODS OR ENGINEERING AEROSPACE OR CONSTRUCTION BUILDING TECHNOLOGY OR REMOTE SENSING OR ENGINEERING OCEAN OR OPERATIONS RESEARCH MANAGEMENT SCIENCE OR ACOUSTICS OR COMPUTER SCIENCE INTERDISCIPLINARY APPLICATIONS OR MARINE FRESHWATER BIOLOGY OR ENGINEERING INDUSTRIAL OR ZOOLOGY OR PHYSICS MATHEMATICAL OR MATHEMATICS APPLIED ) AND [excluding] Web of Science Categories=( ASTRONOMY ASTROPHYSICS OR GEOSCIENCES MULTIDISCIPLINARY ) Databases=SCI-EXPANDED Timespan=1991-2010Lemmatization=Off




16

Supplementary Table 2 – Countries included for each technology

BiofuelsEnergy

EfficiencySolar

Energy WindAustria X X X XCanada X X X XDenmark X X XFrance X X X XGermany X X X XItaly X X X XJapan X X X XNetherlands X X X XNew Zealand XNorway X X X XPortugal X X XSpain X X X XSweden X X XSwitzerland X X X XUnited Kingdom X X XUnited States X X X XTOTAL 15 14 14 14




17

Supplementary Table 3 – Top 10 publication sources, by technology

Biofuels Energy EfficiencyUnited States 4428.62 United States 3753.56Peoples R China 1817.09 Peoples R China 1932.32India 1279.08 Japan 1308.66Brazil 949.75 South Korea 839.05Turkey 793.54 United Kingdom 834.05Japan 761.61 Germany 719.78United Kingdom 749.06 Canada 634.77Canada 735.17 Italy 613.12Germany 735.05 Taiwan 545.87Spain 714.40 France 477.72

Solar Energy WindUnited States 6323.95 United States 916.47Peoples R China 5044.19 United Kingdom 571.13Japan 4314.37 Denmark 337.15Germany 3525.28 Germany 290.77South Korea 2415.65 Spain 268.43India 2123.81 Peoples R China 255.60Taiwan 1506.67 Canada 251.87United Kingdom 1409.72 Japan 222.81France 1219.91 Greece 200.63Spain 1187.10 Turkey 197.62




18

Supplementary Table 4– Descriptive Statistics for R&D and Publicationstechnology variable N mean sd min maxbiofuels R&D 315 25.48 86.02 0.00 1186.90

weighted publications 315 35.06 96.38 0.00 1138.25energy efficiency R&D 294 102.31 216.09 0.01 2149.17

weighted publications 294 29.94 55.95 0.00 493.99solar energy R&D 294 33.49 51.17 0.00 401.70

weighted publications 294 74.66 128.47 0.00 1208.84wind R&D 294 10.69 17.54 0.00 187.19

weighted publications 294 11.83 20.03 0.00 188.20Total R&D 1197 42.69 123.79 0.00 2149.17

weighted publications 1197 37.82 88.68 0.00 1208.84NOTES: R&D in millions of 2010 US $




19

Supplementary Table 5 – Other Variables and data sourcesV

aria

ble

Sour

ceB

iofu

els

Ene

rgy

Eff

icie

ncy

Sola

rW

ind

Polic

y Va

riab

les

Gas

olin

e ta

xes (

2010

US

$ pe

r lite

r)IE

AX

% re

new

able

s req

uire

d by

RPS

OEC

D-E

PAU

XX

Feed

-in ta

riff,

sola

r PV

(201

0 U

S $

per k

Wh)

OEC

D-E

PAU

XFe

ed-in

tarif

f, w

ind

(201

0 U

S $

per k

Wh)

OEC

D-E

PAU

XC

ontr

ol V

aria

bles

ln(p

er c

apita

GD

P) (2

005

US$

)O

ECD

XX

XX

Gas

olin

e pr

ices

, w/o

ut ta

xes (

2010

US

$ pe

r lite

r)IE

AX

Gas

olin

e pr

ices

, inc

. tax

es (2

010

US

$ pe

r lite

r)IE

AX

Hou

seho

ld e

lect

ricity

pric

e (2

010

US$

per

MW

h)IE

AX

Cru

de o

il pr

oved

rese

rves

per

cap

ita (m

illio

n ba

rrel

s)EI

A/O

ECD

aX

XN

atur

al g

as p

rove

d re

serv

es p

er c

apita

(bill

ion

cu. f

t.)EI

A/O

ECD

aX

Coa

l pro

duct

ion

per c

apita

(sho

rt to

ns)

EIA

/OEC

Da

X%

ele

ctric

ity fr

om h

ydro

pow

erEI

AX

X%

ele

ctric

ity fr

om n

ucle

ar

EIA

XX

grow

th ra

te o

f ele

ctric

ity c

onsu

mpt

ion

EIA

XX

Inst

rum

ents

Dum

my:

is e

xecu

tive

bran

ch p

arty

orie

ntat

ion

right

-win

gD

PIX

XX

XD

umm

y: is

exe

cutiv

e br

anch

par

ty o

rient

atio

n ce

nter

DPI

XX

XX

Tax

reve

nue

(exc

ludi

ng so

cial

secu

rity)

, % o

f GD

PO

ECD

XX

XX

Gen

eral

gov

ernm

ent c

onsu

mpt

ion

expe

nditu

re, %

of G

DP

OEC

DX

XX

XG

over

nmen

t ene

rgy

R&

D: b

iofu

els

IEA

XX

Gov

ernm

ent e

nerg

y R

&D

: ene

rgy

effic

ienc

yIE

AX

XG

over

nmen

t ene

rgy

R&

D: e

nerg

y st

orag

eIE

AX

XG

over

nmen

t ene

rgy

R&

D: n

ucle

ar fu

sion

IEA

XG

over

nmen

t ene

rgy

R&

D: s

olar

ene

rgy

IEA

Gov

ernm

ent e

nerg

y R

&D

: win

dIE

AX

XD

PI: D

atab

ase

of P

oliti

cal I

nstit

utio

ns 2

012;

1EI

A: U

S En

ergy

Info

rmat

ion

Adm

inis

tratio

n; IE

A: I

nter

natio

nal E

nerg

y A

genc

y En

ergy

Pr

ices

and

Tax

es D

atab

ase;

OEC

D:O

ECD

Stat

; OEC

D-E

PAU

: OEC

D R

enew

able

Ene

rgy

Polic

y D

atab

ase,

2 an

upda

te a

nd e

xten

sion

of

the

data

set o

rigin

ally

use

d in

John

ston

e et

al.3

a : Res

erve

dat

a fr

om E

nerg

y In

form

atio

n A

dmin

istra

tion;

pop

ulat

ion

data

to c

alcu

late

per

cap

ita v

alue

s fro

m O

ECD

Stat




20

Supplementary Table 6 – AIC statistic for alternative number of years of lagged R&DBi

ofue

lsEn

ergy

Effi

cien

cyFD

IVP

DL

IVFD

IVP

DL

IVal

l con

trols

sing

le y

ear

all c

ontro

lssi

ngle

yea

ral

l con

trols

sing

le y

ear

all c

ontro

lssi

ngle

yea

rA

IC la

g 1:

2811

.928

06.3

AIC

lag

1:21

8221

71.6

AIC

lag

2:24

06.5

2400

.5A

IC la

g 2:

2059

.720

45A

IC la

g 3:

2408

2402

.4A

IC la

g 3:

2070

.920

46.8

AIC

lag

4:24

01.2

2401

.924

01.2

2435

.8A

IC la

g 4:

2079

.420

47.3

2076

.420

47.3

AIC

lag

5:23

97.2

2403

.824

25.6

2434

.8A

IC la

g 5:

2083

.420

44.3

2071

.520

53.6

AIC

lag

6:24

01.6

2399

2425

.624

28.4

AIC

lag

6:20

9120

46.3

2086

.320

53.9

AIC

lag

7:23

94.1

2400

.724

21.9

2428

.5A

IC la

g 7:

2097

.320

48.5

2096

.520

62.1

AIC

lag

8:23

78.3

2400

.224

13.7

2427

.5A

IC la

g 8:

2105

.920

48.7

2093

.320

69.4

AIC

lag

9:23

5623

72.6

2388

.824

17.5

AIC

lag

9:21

1420

47.8

2114

.120

67.5

AIC

lag

10:

2351

.323

56.9

2389

.823

97.2

AIC

lag

10:

2110

2038

.121

30.6

2062

.1A

IC la

g 11

:23

51.5

2352

.324

08.2

2420

.1A

IC la

g 11

:21

14.1

2038

.821

33.5

2063

.1

Sola

r Ene

rgy

Win

d En

ergy

FDIV

PD

L IV

FDIV

PD

L IV

all c

ontro

lssi

ngle

yea

ral

l con

trols

sing

le y

ear

all c

ontro

lssi

ngle

yea

ral

l con

trols

sing

le y

ear

AIC

lag

1:27

42.8

2732

.5A

IC la

g 1:

1839

.518

44.9

AIC

lag

2:27

04.1

2688

.2A

IC la

g 2:

1738

.717

42.5

AIC

lag

3:27

14.2

2689

.9A

IC la

g 3:

1748

.117

44A

IC la

g 4:

2672

.926

41.2

2652

.626

32.9

AIC

lag

4:17

50.8

1745

1764

.317

45.0

AIC

lag

5:26

74.6

2643

.226

57.3

2636

.6A

IC la

g 5:

1756

.517

46.9

1771

.817

58.3

AIC

lag

6:26

76.4

2640

.426

54.2

2625

.5A

IC la

g 6:

1763

.317

47.8

1773

.217

65.3

AIC

lag

7:26

76.8

2641

2644

.726

24.6

AIC

lag

7:17

66.5

1737

.317

70.6

1760

.5A

IC la

g 8:

2682

.426

40.3

2638

.926

23.7

AIC

lag

8:17

77.1

1739

.317

88.2

1767

.7A

IC la

g 9:

2689

.126

36.9

2631

.826

26.0

AIC

lag

9:17

47.7

1740

.217

58.6

1768

.7A

IC la

g 10

:27

60.2

2639

.426

23.8

2626

.0A

IC la

g 10

:17

46.3

1737

.717

41.3

1768

.7

a: 2

nd d

egre

e po

lyno

mia

l; b:

3rd

deg

ree

poly

nom

ial;

c: 4

th d

egre

e po

lyno

mia

l




21

Supplementary Table 7 – First-stage energy R&D results (Instruments and lagged own energy R&D)

Biofuels Eng. Eff. Solar WindR&D: Biofuels 0.170*** 0.0670**

(0.0287) (0.0276)R&D: Energy Efficiency 0.481*** 0.0546***

(0.0892) (0.0184)R&D: Energy Storage 0.578 2.351***

(0.463) (0.652)R&D: Nuclear Fusion -0.00596

(0.120)R&D: Wind 0.410 4.561***

(0.408) (0.937)Political Leadership: Right -0.526 -12.51 -2.991 0.648

(3.849) (11.14) (3.056) (1.391)Political Leadership: Center -7.513 -4.975 -4.863 -4.331

(14.87) (16.87) (6.982) (6.353)Govt. Consumption (% GDP) -5.201 -7.246 -0.250 0.794

(3.776) (6.549) (2.565) (0.956)Tax Revenue (% GDP) -1.281 0.476 -1.610 -0.317

(1.975) (3.788) (1.240) (0.663)RD(t-1) 0.0220 0.213*** -0.00981 -0.286***

(0.0974) (0.0706) (0.0831) (0.0622)RD(t-2) -0.0586 -0.195*** 0.247** -0.0789

(0.0901) (0.0480) (0.125) (0.105)RD(t-3) 0.0644 -0.131 -0.214 -0.209*

(0.259) (0.146) (0.134) (0.121)RD(t-4) -0.195 0.101 -0.0371 0.0729

(0.352) (0.186) (0.125) (0.113)RD(t-5) -0.162 -0.114 -0.239 -0.163

(0.274) (0.135) (0.232) (0.122)RD(t-6) 0.0273 0.199 0.146 -0.0997

(0.436) (0.205) (0.199) (0.0974)RD(t-7) 0.0663 0.291* -0.133

(0.231) (0.148) (0.105)RD(t-8) 0.556 -0.372*

(0.456) (0.217)RD(t-9) 0.792* -0.181

(0.417) (0.146)RD(t-10) -0.191 -0.0391

(0.550) (0.221)N 300 280 280 280F-stat all instruments 32.51 17.50 11.40 38.49Standard errors in parenthesesAll models use robust standard errors with correction for autocorrelation.* p<0.1, ** p<0.05, *** p<0.01




22

Supplementary Table 8 -- First-stage energy R&D results (controls)

Biofuels Eng. Eff. Solar WindR&D(t-1) through (t-10) 0.922 -0.229 -0.107 -0.897**

(1.050) (0.379) (0.405) (0.372)gas tax -49.517

(56.72)FIT dummies

FIT wind -5.672(72.61)

FIT pv -38.819(34.18)

FIT geo

REC dummy

REC levels 3.13 0.292(4.041) (2.384)

% hydro -2.818** -0.351(1.356) (1.444)

% nuclear 0.046 -1.358(1.936) (1.241)

lnGDP -94.652 375.008 -155.17 25.817(180.2) (566.3) (123.0) (61.52)

grow_elec -2.685 -2.315(3.593) (1.955)

gas price 98.727(131.3)

gas price no taxes 49.743(153.7)

oil per capita -17565.6 -108127(24391.6) (89925.4)

gas per capita 19784.34(14992.1)

coal per capita 44.093(47.62)

electric price -0.136(0.749)

N 300 280 280 280Standard errors in parenthesesAll models use robust standard errors with correction for autocorrelation.* p<0.1, ** p<0.05, *** p<0.01




23

Supplementary Table 9 – First-differenced panel regression results: Government R&DIV

exog

IVex

ogIV

exog

IVex

ogR

D0.

152*

**0.

152*

**

0.

0184

***

0.01

13**

-0

.026

70.

372*

**0.

0983

**0.

124*

**(0

.015

3)(0

.015

1)

(0.0

0648

)

(0.

0052

4)

(0.2

76)

(0.0

839)

(0.0

440)

(0.0

246)

RD

(t-1)

0.37

6***

0.37

6***

0.04

57**

*

0.

0443

***

0.60

1***

0.66

3***

0.19

7***

0.21

2***

(0.0

259)

(0.0

245)

(0

.006

75)

(

0.00

705)

(0

.167

)(0

.150

)(0

.052

0)(0

.046

2)R

D(t-

2)

0.

429*

**0.

429*

**

0.

0718

***

0.06

97**

*0.

483*

*0.

331*

0.33

5***

0.34

8***

(0.0

196)

(0.0

196)

(0

.008

78)

(

0.00

845)

(0

.246

)(0

.200

)(0

.054

7)(0

.054

9)R

D(t-

3)

0.

0071

90.

0072

1

-0.

0136

-0.

0185

0.

154

0.31

2**

0.03

130.

0525

(0.0

719)

(0.0

720)

(0

.029

6)

(0

.029

4)

(0.1

44)

(0.1

27)

(0.0

580)

(0.0

587)

RD

(t-4)

0.17

6**

0.17

6**

0.01

48

0.01

90

0.44

7**

0.60

1***

0.02

630.

0287

(0.0

891)

(0.0

891)

(0

.025

3)

(0

.025

3)

(0.1

91)

(0.2

25)

(0.0

585)

(0.0

575)

RD

(t-5)

-0.1

01-0

.100

-

0.03

06

-

0.03

24

0.03

500.

116

0.01

300.

0216

(0.1

09)

(0.1

09)

(0

.025

5)

(0

.025

3)

(0.2

43)

(0.2

25)

(0.0

532)

(0.0

523)

RD

(t-6)

0.09

400.

0942

0.01

55

0

.008

91

0.38

5**

0.34

7*0.

0618

0.06

39(0

.104

)(0

.104

)

(0.0

174)

(0.0

165)

(0

.194

)(0

.186

)(0

.051

9)(0

.051

5)R

D(t-

7)

0.

0585

0.05

88

0.0

0067

3

0

.003

61

0.12

1***

0.12

7***

(0.0

965)

(0.0

956)

(0

.022

9)

(0

.022

2)

(0.0

451)

(0.0

436)

RD

(t-8)

0.08

840.

0882

-

0.03

29

-

0.03

31

(0.1

18)

(0.1

16)

(0

.025

4)

(0

.024

9)

RD

(t-9)

0.55

8***

0.55

8***

0.03

85

0.04

08

(0.1

60)

(0.1

62)

(0

.026

7)

(0

.026

6)

RD

(t-10

)

0.31

8**

0.31

9**

0.06

01**

*

0.

0616

***

(0.1

52)

(0.1

51)

(0

.021

0)

(0

.021

2)

Cum

ulat

ive e

ffect

s:

R

&D

2.

155*

**2.

158*

**

0.1

88**

*

0.17

5***

2.07

8***

2.74

2***

0.88

3***

0.97

7***

(0.3

48)

(0.3

53)

(0

.052

1)

(0.0

511)

(0.5

99)

(0.5

73)

(0.2

25)

(0.2

02)

N30

030

0

280

280

280

280

280

280

AIC

23

51.3

2351

.3

21

10.0

21

08.3

2676

.426

48.1

1766

.517

65.3

BIC

27

03.2

2703

.2

25

35.3

25

33.5

2999

.929

71.5

2115

.521

14.3

EU

x y

ear c

hi-s

quar

ed23

.11

23.0

4

2

0.00

19

.94

26.3

227

.84

34.6

634

.54

EU

x y

ear p

-val

ue0.

284

0.28

7

0

.458

0.

462

0.15

50.

113

0.02

200.

0227

F R

D 1

st s

tage

32.5

1

1

7.50

11

.40

38.4

9H

anse

n J

p-va

lue

0.

293

0.1

51

0.49

10.

502

End

ogen

eity

test

p-v

alue

0.97

9

0

.687

0.

0459

0.25

8S

tand

ard

erro

rs in

par

enth

eses

. A

ll m

odel

s us

e ro

bust

sta

ndar

d er

rors

with

cor

rect

ion

for a

utoc

orre

latio

n.*:

sig

nific

ant a

t 10%

leve

l. **

: sig

nific

ant a

t 5%

leve

l. *

**: S

igni

fican

t at 1

% le

vel.

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




24

Supplementary Table 10 – First differenced panel regression results: ControlsIV

exog

IVex

ogIV

exog

IVex

ogC

umul

ative

effe

cts:

R&

D

2.15

5***

2.15

8***

0

.188

***

0.

175*

**2.

078*

**2.

742*

**0.

883*

**0.

977*

**(0

.348

)(0

.353

)

(0.0

521)

(0

.051

1)(0

.599

)(0

.573

)(0

.225

)(0

.202

) ln

GD

P

-6

.353

-6.5

49

54

.831

60

.706

-290

.964

*-2

17.0

63-4

7.68

5-4

8.34

7(7

4.85

)(7

4.97

)

(73

.35)

(7

2.62

)(1

74.4

)(1

62.4

)(3

3.00

)(3

3.09

) g

as p

rice

no ta

xes

2

40.4

49**

*24

0.23

***

(81.

66)

(81.

56)

gas

tax

-3

.763

-3.7

52(2

5.22

)(2

5.21

) g

as p

rice

-16

.975

-1

7.86

5

(19

.51)

(1

9.66

) o

il pe

r cap

ita

7211

.473

7235

.793

-8

090.

286

-9

278.

706

(101

75.8

)(1

0120

.7)

(1

7340

.3)

(1

7305

.9)

gas

per

cap

ita

-

1388

.5

-124

5.72

2

(219

3.2)

(2

181.

3) c

oal p

er c

apita

-11.

051*

-1

2.21

*

(6.

395)

(6

.566

) e

lect

ric p

rice

0.

100

0.10

7

(0.

121)

(0

.122

) g

row

_ele

c

-5.2

58-3

.77

-0.8

72-0

.802

(5.6

88)

(5.0

39)

(0.9

20)

(0.9

29)

% h

ydro

0.62

91.

627

1.04

30.

98(2

.065

)(1

.788

)(0

.718

)(0

.737

) %

nuc

lear

-4

.819

*-4

.592

*-0

.499

-0.4

21(2

.493

)(2

.454

)(0

.678

)(0

.681

) F

IT w

ind

9.

827

12.0

61(2

5.86

)(2

6.07

) F

IT p

v

18

.874

39.5

23(6

0.28

)(5

3.99

) R

EC

leve

ls

15.0

92**

13.5

24**

2.13

32.

102

(7.2

34)

(6.3

52)

(1.3

76)

(1.3

52)

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




25

Supplementary Table 11 – Effect of EU-specific year effects: Government R&DE

U x

yea

rno

EU

EU

x y

ear

no E

UE

U x

yea

rno

EU

EU

x y

ear

no E

UR

D0.

152*

**

0

.159

***

0.01

84**

*

0.

0187

***

-0.0

267

0

.007

68

0.09

83**

0.09

56**

(0

.015

3)

(0.0

178)

(0.0

0648

)

(0.

0068

8)

(0.2

76)

(

0.31

3)

(0.0

440)

(0

.046

9)

RD

(t-1)

0.37

6***

0.3

61**

*

0.

0457

***

0.04

40**

*0.

601*

**

0

.609

***

0.19

7***

0.1

96**

*(0

.025

9)

(0.0

296)

(0.0

0675

)

(0.

0076

5)

(0.1

67)

(

0.21

6)

(0.0

520)

(0

.057

1)

RD

(t-2)

0.42

9***

0.3

97**

*

0.

0718

***

0.07

29**

*0.

483*

*

0

.566

*

0.33

5***

0.3

38**

*(0

.019

6)

(0.0

229)

(0.0

0878

)

(0.

0095

4)

(0.2

46)

(

0.30

1)

(0.0

547)

(0

.063

5)

RD

(t-3)

0.00

719

0.07

26

-

0.01

36

-

0.01

60

0.15

4

0

.141

0.

0313

0.04

12

(0.0

719)

(0

.078

9)

(0

.029

6)

(0

.032

3)

(0.1

44)

(

0.17

2)

(0.0

580)

(0

.061

3)

RD

(t-4)

0.17

6**

0.2

94**

0.

0148

0.

0111

0.

447*

*

0

.390

*

0.02

63

0.

0222

(0

.089

1)

(0.

121)

(0.0

253)

(0.0

272)

(0

.191

)

(0.

220)

(0

.058

5)

(0.0

662)

R

D(t-

5)

-0

.101

-

0.02

70

-

0.03

06

-

0.01

34

0.03

50

0

.136

0.

0130

0

.006

34

(0.1

09)

(

0.11

8)

(0

.025

5)

(0

.028

5)

(0.2

43)

(

0.30

3)

(0.0

532)

(0

.059

4)

RD

(t-6)

0.09

40

0

.139

0.

0155

0.

0140

0.

385*

*

0

.374

*

0.06

18

0.

0942

(0

.104

)

(0.

127)

(0.0

174)

(0.0

180)

(0

.194

)

(0.

219)

(0

.051

9)

(0.0

596)

R

D(t-

7)

0.

0585

0.05

87

0

.000

673

0.

0142

0.

121*

**

0

.145

***

(0.0

965)

(

0.12

1)

(0

.022

9)

(0

.024

7)

(0.0

451)

(0

.051

0)

RD

(t-8)

0.08

84

0.0

0444

-0.

0329

-0.

0356

(0

.118

)

(0.

147)

(0.0

254)

(0.0

279)

R

D(t-

9)

0.

558*

**

0

.559

***

0.03

85

0.04

28

(0.1

60)

(

0.20

9)

(0

.026

7)

(0

.031

7)

RD

(t-10

)

0.31

8**

0.3

65*

0.

0601

***

0.06

49**

*(0

.152

)

(0.

214)

(0.0

210)

(0.0

191)

C

umul

ative

effe

cts:

R&

D

2.15

5***

2

.383

***

0.1

88**

*

0

.218

***

2.

078*

**

2.2

23**

*

0.88

3***

0

.939

***

(0

.348

)

(0.

444)

(0.0

521)

(0.0

534)

(0

.599

)

(0.

727)

(0

.225

)

(0.

246)

N

300

30

0

280

28

028

0

280

280

28

0A

IC

2351

.3

23

67.0

21

10.0

21

02.8

26

76.4

2682

.1

1766

.5

17

60.1

B

IC

2703

.2

26

44.8

25

35.3

24

29.9

29

99.9

2932

.9

2115

.5

20

36.3

E

U x

yea

r p-v

alue

0.28

4

0

.458

0.

155

0.02

20F

RD

1st

sta

ge

32

.51

29.

82

17.

50

19.

95

11.4

0

1

3.44

38

.49

32.

07

Han

sen

J p-

valu

e

0.29

3

0

.411

0

.151

0

.109

0.

491

0.6

32

0.50

2

0

.540

E

ndog

enei

ty te

st p

-val

ue

0.

979

0.7

37

0.6

87

0.5

95

0.04

59

0

.124

0.

258

0.1

92

Sta

ndar

d er

rors

in p

aren

thes

es.

All

mod

els

use

robu

st s

tand

ard

erro

rs w

ith c

orre

ctio

n fo

r aut

ocor

rela

tion.

*: s

igni

fican

t at 1

0% le

vel.

**: s

igni

fican

t at 5

% le

vel.

***

: Sig

nific

ant a

t 1%

leve

l.

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




26

Supplementary Table 12 – Effect of EU-specific year effects: ControlsE

U x

yea

rno

EU

EU

x y

ear

no E

UE

U x

yea

rno

EU

EU

x y

ear

no E

UC

umul

ative

effe

cts:

R&

D

2.15

5***

2

.383

***

0.1

88**

*

0

.218

***

2.

078*

**

2.2

23**

*

0.88

3***

0

.939

***

(0

.348

)

(0.

444)

(0.0

521)

(0.0

534)

(0

.599

)

(0.

727)

(0

.225

)

(0.

246)

ln

GD

P

-6

.353

-41.

66

54.8

31

65.0

24

-290

.964

*

-231

.318

-4

7.68

5

-40

.879

(7

4.85

)

(87

.24)

(73

.35)

(71

.26)

(1

74.4

)

(17

5.7)

(3

3.00

)

(34

.85)

g

as p

rice

no ta

xes

2

40.4

49**

* 2

93.5

45**

*

(81.

66)

(

94.1

9)

gas

tax

-3

.763

-8.8

2

(25.

22)

(

28.1

4)

gas

pric

e

-

16.9

75

-

10.4

72

(

19.5

1)

(

21.3

3)

oil

per c

apita

72

11.4

73

583

3.28

9

-8

090.

286

987

9.78

6

(101

75.8

)

(847

3.9)

(173

40.3

)

(13

051.

2)

gas

per

cap

ita

-

1388

.5

-1

346.

34

(2

193.

2)

(2

100.

3)

coa

l per

cap

ita

-1

1.05

1*

-7.5

4

(

6.39

5)

(

7.20

1)

ele

ctric

pric

e

0.10

0

0

.108

(0.

121)

(0.

108)

g

row

_ele

c

-5.2

58

2

.215

-0

.872

-0.0

33

(5.6

88)

(

4.95

9)

(0.9

20)

(

0.88

8)

% h

ydro

0.62

9

1

.102

1.

043

1.2

36

(2.0

65)

(

2.33

5)

(0.7

18)

(

0.79

4)

% n

ucle

ar

-4.8

19*

-1.8

43

-0.4

99

-0

.163

(2

.493

)

(2.

551)

(0

.678

)

(0.

677)

F

IT w

ind

9.

827

16.4

95

(25.

86)

(

27.5

7)

FIT

pv

18.8

74

-70

.981

(6

0.28

)

(66

.22)

R

EC

leve

ls

15.0

92**

7.0

73

2.13

3

1

.753

(7

.234

)

(7.

325)

(1

.376

)

(1.

440)

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




27

Supplementary Table 13 – Polynomial distributed lag regression results: Government R&D

IVex

ogIV

exog

IVex

ogIV

exog

RD

0.14

35**

*0.

1449

***

0.02

45**

*0.

0100

0.26

78**

0.39

09**

*0.

1085

*0.

1276

***

(0.0

123)

(0.0

143)

(0.0

064)

(0.0

073)

(0.1

228)

(0.0

852)

(0.0

568)

(0.0

323)

RD

(t-1)

0.31

21**

*0.

3127

***

0.03

41**

*0.

0410

***

0.38

72**

*0.

4460

***

0.25

07**

*0.

2647

***

(0.0

278)

(0.0

283)

(0.0

083)

(0.0

053)

(0.1

041)

(0.1

013)

(0.0

451)

(0.0

411)

RD

(t-2)

0.38

22**

*0.

3826

***

0.04

07**

*0.

0506

***

0.45

32**

*0.

4692

***

0.26

58**

*0.

2764

***

(0.0

181)

(0.0

181)

(0.0

147)

(0.0

084)

(0.1

431)

(0.1

405)

(0.0

514)

(0.0

551)

RD

(t-3)

0.37

68**

*0.

3774

***

0.04

44**

0.04

47**

*0.

4655

***

0.46

04**

*0.

2021

***

0.21

04**

*(0

.036

6)(0

.036

5)(0

.019

2)(0

.010

0)(0

.162

0)(0

.154

9)(0

.055

5)(0

.060

4)R

D(t-

4)

0.

3210

***

0.32

19**

*0.

0451

**0.

0291

***

0.42

43**

*0.

4198

***

0.10

78*

0.11

47*

(0.0

614)

(0.0

615)

(0.0

212)

(0.0

104)

(0.1

518)

(0.1

461)

(0.0

552)

(0.0

594)

RD

(t-5)

0.24

15**

*0.

2426

***

0.04

29**

0.00

950.

3295

**0.

3472

**0.

0312

0.03

70(0

.072

9)(0

.073

6)(0

.020

8)(0

.010

4)(0

.135

5)(0

.141

9)(0

.049

7)(0

.052

3)R

D(t-

6)

0.

1669

**0.

1681

**0.

0378

**-0

.008

30.

1811

0.24

280.

0204

0.02

52(0

.072

0)(0

.073

7)(0

.018

0)(0

.010

3)(0

.179

7)(0

.196

6)(0

.039

7)(0

.039

7)R

D(t-

7)

0.

1274

*0.

1287

*0.

0297

**-0

.018

3*0.

1237

**0.

1271

**(0

.071

7)(0

.074

4)(0

.013

7)(0

.010

1)(0

.051

7)(0

.049

7)R

D(t-

8)

0.

1553

*0.

1564

*0.

0186

*-0

.014

9(0

.085

5)(0

.088

4)(0

.010

9)(0

.010

2)R

D(t-

9)

0.

2843

**0.

2854

**0.

0046

0.00

78(0

.117

5)(0

.119

5)(0

.016

2)(0

.013

4)R

D(t-

10)

0.

5502

***

0.55

15**

*-0

.012

30.

0556

**(0

.197

1)(0

.197

8)(0

.028

4)(0

.023

0)C

umul

ative

effe

cts:

R&

D

3.06

12**

*3.

0723

***

0.31

02**

*0.

2067

***

2.50

85**

*2.

7764

***

1.11

01**

*1.

1830

***

(0.4

413)

(0.4

611)

(0.1

001)

(0.0

633)

(0.6

621)

(0.7

313)

(0.2

765)

(0.2

792)

N30

030

028

028

028

028

028

028

0A

IC

2389

.823

89.8

2085

.220

84.1

2641

.126

37.1

1770

.617

70.0

BIC

26

30.5

2630

.523

32.3

2331

.328

62.8

2858

.820

17.8

2017

.1S

tand

ard

erro

rs in

par

enth

eses

. A

ll m

odel

s us

e ro

bust

sta

ndar

d er

rors

with

cor

rect

ion

for a

utoc

orre

latio

n.*:

sig

nific

ant a

t 10%

leve

l. **

: sig

nific

ant a

t 5%

leve

l. *

**: S

igni

fican

t at 1

% le

vel.

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




28

Supplementary Table 14 – Polynomial distributed lag regression results: ControlsIV

exog

IVex

ogIV

exog

IVex

ogC

umul

ative

effe

cts:

R&

D

3.06

12**

*3.

0723

***

0.31

02**

*0.

2067

***

2.50

85**

*2.

7764

***

1.11

01**

*1.

1830

***

(0.4

413)

(0.4

611)

(0.1

001)

(0.0

633)

(0.6

621)

(0.7

313)

(0.2

765)

(0.2

792)

lnG

DP

-2.1

057

-2.8

775

71.1

088

51.4

496

-1.9

e+02

-1.7

e+02

-41.

2015

-41.

6738

(82.

2929

)(8

1.76

42)

(87.

4775

)(6

9.23

78)

(169

.493

9)(1

68.1

542)

(33.

6773

)(3

3.87

59)

gas

pric

e no

taxe

s

221.

8810

***

221.

2094

***

(84.

4800

)(8

3.84

67)

gas

tax

-3

.194

1-3

.139

3(2

7.23

91)

(27.

1973

) g

as p

rice

-4

.349

6-6

.853

9(2

4.26

36)

(21.

2315

) o

il pe

r cap

ita

7.1e

+03

7.1e

+03

-1.1

e+04

-1.2

e+04

( 1.1

e+04

)( 1

.1e+

04)

( 8.2

e+03

)( 8

.4e+

03)

gas

per

cap

ita

-4.0

e+03

-1.5

e+03

( 2.5

e+03

)( 2

.2e+

03)

coa

l per

cap

ita

-16.

5876

**-1

2.09

52**

(6.5

131)

(5.2

584)

ele

ctric

pric

e

0.06

940.

0774

(0.1

283)

(0.1

238)

gro

w_e

lec

-3

.287

8-2

.704

7-0

.771

7-0

.716

7(5

.142

4)(4

.954

5)(1

.016

9)(1

.002

6) %

hyd

ro

-0

.074

30.

0766

0.44

730.

4144

(1.6

134)

(1.5

613)

(0.8

803)

(0.8

923)

% n

ucle

ar

-3.6

867

-3.6

852

-0.0

177

0.03

23(2

.575

5)(2

.509

7)(0

.755

3)(0

.768

5) F

IT w

ind

25

.060

025

.836

9(2

6.42

34)

(26.

5270

) F

IT p

v

55

.477

955

.773

9(6

0.09

10)

(59.

4518

) R

EC

leve

ls

14.4

206*

*13

.540

7*2.

6813

*2.

6451

*(7

.239

4)(6

.929

3)(1

.556

0)(1

.526

8)S

tand

ard

erro

rs in

par

enth

eses

. A

ll m

odel

s us

e ro

bust

sta

ndar

d er

rors

with

cor

rect

ion

for a

utoc

orre

latio

n.*:

sig

nific

ant a

t 10%

leve

l. **

: sig

nific

ant a

t 5%

leve

l. *

**: S

igni

fican

t at 1

% le

vel.

Bio

fuel

sE

nerg

y E

ffici

ency

Sol

ar E

nerg

yW

ind

Ene

rgy




29

Supplementary Table 15 – Time to first patent citation hazard regression

full sample articles from 2000 or laterBiofuels Solar Wind Biofuels Solar Wind

citation lag 0.706*** 0.333*** 0.531*** 1.359*** 0.691*** 0.973**(0.142) (0.0500) (0.157) (0.326) (0.109) (0.480)

(citation lag)^2 -0.0249*** -0.0167*** -0.0191*** -0.0682*** -0.0412*** -0.0562(0.00754) (0.00331) (0.00731) (0.0231) (0.00866) (0.0377)

Multiple country dummy -0.0301 -0.0294 -1.234** -0.0478 -0.0821 -1.775*(0.282) (0.123) (0.626) (0.322) (0.151) (1.058)

Constant -8.987*** -7.500*** -7.353*** -11.22*** -8.621*** -8.283***(0.948) (0.597) (1.096) (1.324) (0.659) (1.655)

N 55790 164957 21895 35568 95725 13670AIC 1407.3 4306.9 831.2 881.6 2444.9 487.3BIC 3827.1 6930.5 2709.8 2170.4 3780.1 1525.5log likelihood -432.6 -1891.5 -180.6 -288.8 -1081.5 -105.6* p<0.10, ** p<0.05, *** p<0.01Standard errors in parentheses. All standard errors clustered by article. All regressions include publication year x country fixed effects, with publications from 2001 in the US as the base category.




30

SUPPLEMENTARY NOTES

Supplementary Note 1. Web of Science Keyword Searches

Our publication data were provided as a custom database from Thomson Reuters, created

using keyword searches in Web of Science developed by the researchers in consultation with

experts at Thomson Reuters. As noted in the text, we devised searches that would be narrow, so

as to avoid irrelevant articles, as such articles would respond differently to alternative energy

R&D trends and thus bias our results downward. Thus, for each category, we devised keyword

searches that are refined by focusing on specific subject categories in the Web of Science

database. The searches used for each technology are listed in Supplementary Table 1.

Supplementary Note 2. Data Description

Our data set combines the publication data provided by Thomson Reuters with publicly

available energy R&D data from the International Energy Agency. Additional publicly available

data are used as controls. As the R&D data series are incomplete for some countries, the set of

countries included for each technology depends on both the country actively publishing in that

area and energy R&D data being available. R&D data go back as far as 1974 for some countries.

Using a starting date of 1992 for the publication data, we select countries with at least 10 years

of lagged R&D data for a given technology. Supplementary Table 2 lists the countries included

in the regressions for each technology. Supplementary Table 3 lists the top 10 publishing

countries for each technology. While the top 10 includes a few emerging economies, such as

China, our final data set includes most of these top publishing countries.

Supplementary Table 4 provides descriptive statistics for both R&D and the weighted

publications. R&D data are shown in millions of 2010 dollars. Evaluated at the means, for most




31

technologies we see a bit less than one million dollars of R&D spending per publication. The

exception is energy efficiency, for which a little over $3 million of R&D is spent per publication.

While this may indicate that energy efficiency R&D is less productive than other R&D spending,

it may also indicate that our keyword searches identify a lower percentage of relevant energy

efficiency articles than for other technologies.

Figure 1 in the main text shows the aggregate trends in both energy R&D and

publications for all countries in each regression. Here, Supplementary Figures 1-4 illustrate

these trends by technology for each country. Note that there is variation both across times and

across countries, which is important for separately identifying the effect of energy R&D, as

opposed to simply finding that publications increase due to increased global attention to climate

change. For example, prior to the 2009 stimulus, the U.S. experiences two peaks in wind energy

R&D – one after the 1970s energy crises, and another in the mid-1990s. In contrast, Denmark

and the Netherlands have relatively flat R&D budgets, with the exception of the Netherlands in

1985. Note also that countries also choose to emphasize different technologies. For example, in

2010 the United States spent over ten times as much on wind energy as Japan in 2007, but nearly

twice as much over twice as much on biofuels research as on solar energy, whereas Japan spent

more on solar energy R&D than on biofuels. These figures also show the importance of lagged

R&D effects, as within individual countries, R&D spending generally peaks several years before

publication counts. Examples include wind energy in Germany, Italy, and Japan. As suggested

by our empirical results, the importance of lags is particularly evident in the case of energy

efficiency.

Supplementary Table 5 lists the policy variables, controls, and instruments used to

estimate equations (1) – (3) for each technology. The columns indicate which variables are




32

included for each technology. As noted in the methods section, our instruments include two

categories of variables. First, we include instruments for R&D spending on related technologies,

with the specific technologies chosen for each technology indicated in the table.

Overidentification tests were used to verify the validity of all technologies included and to rule

out invalid instruments. Note that any related technologies are not used as instruments. For

example, nuclear R&D is only used as an instrument for biofuels, a transportation-based

technology, but not for any technologies pertaining to electricity. Since the share of nuclear

energy is a control for solar and renewable energy, decisions on nuclear R&D cannot be

considered exogenous for these technologies. Similarly, policies that promote renewable

electricity sources often target both solar and wind energy. Thus, these technologies cannot

serve as instruments for each other.

Second, since all energy R&D decisions are a result of a political process, we also

include instruments that model the political process determining R&D funding.

Overidentification tests assume that at least one instrument is valid. Since all energy R&D

decisions are made through similar political processes, additional instruments that control for the

political pressures that shape R&D decisions help ensure the validity of the overidentification

tests. The policy instruments control for a country’s general proclivity for government spending

and for the political leanings of the government. For example, Baccini and Urpelainen show that

political shifts affect the volatility of public energy R&D.4




33

Supplementary Note 3. Identifying the Lag Length

As noted in the text, we employ several strategies to identify the appropriate lag length.

First, as is common in the literature, the primary criterion is finding the minimum AIC statistic

across a range of models.2 Alternatively, we also calculated the BIC statistic for each model,

which includes a greater penalty for including irrelevant variables. As a result, collinearity

among the lagged R&D values often leads the BIC to recommend fewer lags than the AIC.

However, in the case of collinear lagged R&D values, we can still estimate long-run effects that

are jointly significant, even when individual year coefficients are estimated imprecisely.

Moreover, leaving out relevant lags would lead to omitted variable bias. Thus, I focus on the

findings of the AIC statistic in the discussion that follows. One complication is that our model

includes not only lagged values of the R&D variables, but also of several control variables.

These controls are often individually insignificant. Thus, we also run models including only a

single year of the control and policy variables, to see if this changes the recommended number of

lags. We initially examine models including up to ten lags of R&D. Adding additional lags is

problematic when the model includes renewable energy policy variables as only two countries in

our sample adopted renewable energy certificates before 2003, so that estimates of the lagged

value of the REC variable beyond eight years are unreliable. The first US states to adopt REC

limits do so in 1998, and Italy adopts an REC limit in 2002. Most states first limits appear in the

data in 2003. However, adding up to 11 years of lags is possible for biofuels and energy

efficiency.

In addition, even collinearity among the R&D variables themselves may cause the AIC

statistic to favor smaller lags. Thus, we also run both the full and single control models using a

polynomial distributed lag model (PDL). The PDL models provide similar results to the main




34

first differenced models, but by requiring fewer parameters to estimate multiple lags, they offer

the potential for lower standard errors on the R&D coefficients. Supplementary Table 6 presents

the AIC statistics.

In some cases, these strategies recommend different lag lengths. Thus, we also consider

the cumulative long-run effect of R&D when evaluating the lag length. These long-run effects,

illustrated in Supplementary Figure 5, tend to be similar across the various estimation

techniques, so we focus on the results using first differenced instrumental variables. When the

AIC statistic conflicts across models, we use the pattern of cumulative effects as an additional

guide. In particular, we consider the lag length at which the cumulative effect of government

energy R&D spending levels out, as this is evidence that all relevant lags have been included.

Turning first to biofuels, we see relatively consistent recommendations across models.

Using the AIC criterion we find an optimal lag length of ten years using the full model and

eleven years using a single year of controls. The PDL results suggest an optimal lag length of

nine years in the full model and ten when using a single year of controls. As the AIC is lowest

for the full FD model, which has an optimal lag of ten years, I use a lag length of ten years for

analyzing the biofuels results.

The results for wind demonstrate the potential advantages of the PDL model and for

looking at the results of models with a single year of controls. Using the standard first

differenced model, the AIC suggests an optimal lag length of just two years. Using just a single

year of controls, the AIC suggests using seven years of lagged data. In contrast, the PDL model

suggests ten year lags in the full model and four lags when using a single year of controls. This

difference, however is not caused by adding lags of energy R&D, but rather lags of the tradable

permit policy variable, which becomes large and significant beginning with the ninth lagged




35

variable. However, beyond nine years of lags, this coefficient is identified off of just one

country, as only the United States had renewable credit trading before 2002. Thus, we turn to the

cumulative effect of energy R&D to help sort through the options. In all four models (FD, FD

single control, PDL, PDL single control), the cumulative effect grows significantly between

years 6 and 7 before leveling off, suggesting a lag of seven years (as suggested by the FD model

with a single year of controls) is appropriate. Indeed, the eighth and ninth lags of energy R&D

are insignificant and close to zero. Thus, I use a lag length of seven years for analyzing the wind

energy results.

For solar energy, shorter lags appear to suffice, although the choice using the AIC

statistic varies across models. In the FD model, the AIC statistic is minimized with just four lags

in the full model and with nine lags when using just a single year of controls. In contrast, the

AIC suggests ten years of lags in the PDL model with full controls, and eight when using just a

single year of controls.

Thus, again we turn to the estimated cumulative effects for more information. Here, we

see a large increase in the cumulative effect between years three and four, and again between

years five and six. While the individual coefficient for year t-6 is only significant at the 10

percent level, both the large size of the coefficient, as well as the continued growth in the

cumulative effect through year six, suggests that six years should be included in the solar energy

models. There is little value to adding additional lags of energy R&D in the PDL models.

Indeed, results in year 10 seem to be largely driven by changes in the REC variable, as the sign

of the effect of R&D becomes negative in the model with full controls. In contrast, the

cumulative effect of energy R&D remains level beyond year 6 in the model using a single year

of control variables.




36

The appropriate lag length is more difficult to identify in the case of energy efficiency, as

the effect of energy R&D itself is smaller. In the FD model, the AIC statistic suggest a lag of

two years using the full model but ten years when using a single year of controls. In the PDL

model, the AIC is minimized using a lag of five years using a full set of controls, and four lagged

years with a single year of controls. Given these divergent results, we turn to the cumulative

effect of energy R&D. The cumulative effect levels off after year six before rising again

beginning in year 9. Among models with 9 to 11 lags, the AIC is minimized in the main FD

model when using 10 lags. Thus, I use ten years of lagged R&D when evaluating energy

efficiency.

Supplementary Note 4. Discussion of Instrument Validity

We include two categories of instrumental variables in our model. The first category

includes public R&D spending on other energy technologies. The key assumption is that these

spending levels are related to energy R&D funding for the technology included in each

regression, but do not lead to additional publications for the endogenous technology. For

example, governments may increase funding for biofuels and wind energy at the same time, but

new biofuels R&D funding should not lead to additional wind energy publications. This

assumes that public R&D decisions are part of a common political process. Of course, if that is

the case, then researchers may also change their research strategies after political shifts, with the

expectation that other factors besides R&D funding may make renewable energy research more

(or less desirable). Thus, we also include a set of instruments related to the political process,

such as changes in the political party in power and the share of GDP devoted to government




37

spending. By including these variables, the R&D instruments can be interpreted as the effect of

increases in other R&D spending conditional on changes in the political climate.

The success of this instrumental variable strategy depends on their being common factors

that influence energy R&D funding across multiple sources. Thus, an understanding of the

energy R&D funding process is important. Funding decisions are typically made by national

governments. Although there are some examples of long-term funding decisions, such as

Japan’s New Sunshine program instituted in 1993 to develop solar PV technologies, most

funding decisions are made on an annual basis.6 As a result, political changes may result in

energy R&D funding changes. Examples include reductions in energy funding after President

Reagan took office in the United States in 1981 and after a party change in the leadership of

Denmark in 2002.6 Baccini and Urpelainen provide statistical evidence that changes in party

leadership affect energy R&D volatility.4 Importantly, while common trends such as changes in

global oil prices that our year effects pick up affect R&D funding decisions,7 both the above

evidence and our descriptive data presented in Supplementary section 2 suggest that there is

variation in funding decisions within countries at different times, enabling our instruments to

capture variation in the funding process.

To further assess the ability of our instruments to predict energy R&D funding,

Supplementary Tables 7 and 8 present the first stage regression results for our main models.

Supplementary Table 7 includes the instruments and individual years of lagged energy R&D that

remain in the model, as only current energy R&D is treated as endogenous. Supplementary

Table 8 includes the cumulative effect of the various control variables on current energy R&D.

As noted elsewhere, the F-stat for the joint significance of all instruments is above 10 for all

models, suggesting that the instruments do a good job explaining energy R&D funding. In




38

particular, the level of other types of energy R&D spending do a particularly good job, as most

are significant. These appear to capture most of the variation in energy R&D funding, as none of

the political variables are individually significant. Similarly, with a few exceptions, the current

funding levels of related technologies are even more strongly correlated with energy R&D

funding than lagged R&D levels for the same technology, as shown both by the lack of

significance and generally lower magnitudes for these lagged variables. As in the main models,

the various control variables have little effect on energy R&D spending.

Supplementary Note 5. Detailed Results and Robustness Checks

In this section, we provide tables with detailed regression results for both the main model

specification and several robustness checks. We also discuss tests confirming the validity of our

instrumental variables.

5.1 Main results

Having identified the appropriate lag length, Supplementary Tables 9-10 present results

for the first differenced models for each technology. Supplementary Table 9 shows the results

for government R&D for both individual years and the cumulative effect. Supplementary Table

10 shows the cumulative effects of the various controls included in each model. In each table,

the first column for each technology includes results using instruments for current R&D and the

second includes results assuming all variables are exogenous. All results include robust standard

errors that have been corrected for both heteroskedasticity and autocorrelation. Because of the




39

similarities across models, I focus on the first-differenced panel estimates, which avoid imposing

any structure on the lag process. Results using the PDL models are included in section 5.3

Before turning to technology-specific results, we first discuss general trends across all

models. Public sponsored energy R&D generally leads to increases in publications, although the

magnitude of the effect varies across technologies. Other demand characteristics appear less

important, as most controls are insignificant, both individually and in the cumulative effects over

time. This contrasts with most research focusing on private sector research efforts, where policy

plays an important role.3,8-10 Given the long term nature of basic R&D, it is not necessarily

surprising that demand factors such as policy shocks have less influence on research supported

by the public sector. However, these results also suggest that public R&D funding is not simply

replicating support that would otherwise be provided by the private sector. We discuss

exceptions to these findings in the technology-by-technology results below.

Turning to the quality of our instruments, the Hansen J test reveals that the instruments

are valid in all cases (Supplementary Table 9). We also present the p-value of the endogeneity

test, where the null hypothesis is that the current value of energy R&D is exogenous. We fail to

reject the null hypothesis that energy R&D is exogenous for all technologies except solar energy.

However, as the results are generally unchanged across endogenous and exogenous

specifications, we choose to be conservative and focus on the IV results when presenting the

various robustness checks that follow.

Energy R&D has the largest cumulative effect in biofuels. One million dollars of

additional government R&D support results in slightly more than two new publications over ten

years. One reason for the stronger biofuels effect is the long lag length. While the effect of

public R&D levels out for other technologies after six or seven years, for biofuels we find a




40

strong effect even in years nine and ten. There is a strong contemporary effect, with 0.376

publications being induced in year t-1. The FD results suggest a cyclical pattern, with strong

effects also found in years two and four in the first differenced model before picking up again in

year t-9. One possibility that the long range effects suggest is that public R&D leads to increases

in the supply of scientific personnel and infrastructure devoted to biofuels research. While

verifying such an explanation is beyond the scope of the current data set, such questions about

the long run impact of public R&D are a fruitful avenue for further study. While most individual

controls are insignificant, gasoline prices net of taxes have a large positive impact on

publications. Higher prices make substitutes such as biofuels look more attractive and

researchers seem sensitive to the market potential of biofuels when deciding on research projects.

Also, while the cumulative effect of per capita oil reserves is insignificant, individual year effects

have a significant positive effect in years t-3, t-7, and t-8.

Similar to biofuels, the lagged effect for energy efficiency R&D is long. While the largest

single year effect occurs in year t-2, the magnitudes in year t-10 are similar. Compared to the

other technologies, the magnitude of both the individual year effects and cumulative effects for

energy efficiency are much smaller, with one million of energy efficiency R&D generating just

under 0.2 new publications in the long run. Compared to the other technologies in this study,

energy efficiency covers a wider range of potential applications, including specific equipment for

production, vehicles, and even high technology applications such as computers and server farms.

Thus, the smaller magnitude may be an artifact of our data, which relies on keyword searches to

identify relevant articles and may miss some of these broader applications. However, it is also

possible that, because of the broader nature of energy efficiency research, the costs of generating

new energy efficiency results is larger, as the ability to stand on the shoulders of past researchers




41

may be lower in a more diverse research field. Also, while the cumulative effect of per capita oil

reserves is insignificant, individual year effects have a significant effect in years t-3 through t-7,

with a positive effect in all years except t-5.

Solar energy is the one case in which instrumental variables lead to changes in the results.

We reject the null hypothesis that current energy R&D expenditures are exogenous, as the p-

value of the endogeneity test is just 0.049. The cumulative effect of energy R&D falls from

2.742 to 2.078 when using instruments for current R&D. The differences are largely driven by

changes in the contemporary effect of R&D, which is much larger when exogenous. Recall from

the theory section that current variables not only pick up the supply of research projects, but also

the demand of editors to publish research on a given topic. As this unobserved demand may be

correlated with energy R&D, and that we would expect some time to pass before R&D led to a

new publication, it is likely using instrumental variables helps correcting for such demand

effects. Policy also plays a role in the case of solar, with larger renewable energy targets

inducing more publications.

Finally, for wind energy, the results are consistent across all specifications. In the long-

run, a million dollars of energy R&D leads to approximately one new publication. Unlike

biofuels, the effect is quick, with most new publications occurring within the first three years.

After a brief leveling off, the cumulative effect continues to rise in years six and seven, after

which it levels off again, as shown earlier in Supplementary Figure 5. While none of the

controls have a significant cumulative effect, the immediate lag is significant for both the

percentage of electricity from hydro and nuclear, although unexpectedly positive for both.




42

5.2 Results without European Union fixed effects

To account for unreported energy R&D support from the European Union to its member

countries, the results in the text include separate year fixed effects for EU countries.

Supplementary Table 11 compares the results from the first-differenced instrumental variables

model with and without the EU-specific fixed effects for the R&D coefficients, and

Supplementary Table 12 compares the results for the controls.

Dechezlepretre and Popp report that the EU’s share of renewable energy R&D funding is

small,11 so that unaccounted for EU spending is unlikely to be a major factor in scientific

productivity. For example, EU RD&D investments dedicated to the six technologies covered by

the EU Strategic Energy Technology Plan (wind, solar (photovoltaics and concentrated solar

power), electricity grids, bioenergy, carbon capture and storage, fuel cells and hydrogen and

nuclear fission) accounted for just 11 percent of public R&D spending on these technologies in

2010. Supplementary Tables 11 and 12 confirm that the effect of any omitted EU-wide R&D is

minimal, as the results are virtually the same when excluding the EU-specific year effects. The

cumulative effects of energy R&D range from 6 to 15 percent higher without the EU-specific

year trends, which is consistent with the expected bias, assuming that EU funding follows similar

trends as country-level funding. F-tests for the joint significance of all EU-specific year fixed

effects fail to reject the null hypothesis for all technologies except wind. Nonetheless, to avoid

any bias from unreported EU-wide energy R&D spending, the results in the main text include

EU-specific year fixed effects.




43

5.3 Polynomial Distributed Lag Results

The following tables present results using the polynomial distributed lag (PDL) model.

To accommodate instrumental variables for current energy R&D, I used the following steps:

1. Estimate a first-stage model for R&D using the same number of lagged controls

as in the final PDL regression to be estimated.

2. Obtain predicted energy R&D from this regression.

3. Use the predicted current R&D, along with actual values of lagged R&D, to

construct first differenced data.

4. Use first differenced data to construct the polynomials for estimation.

Polynomials from degree 2-4 were used, as were lags from 4-10 years.

As noted in the text, there are few differences between the PDL results and those simply

including all lags in the model. Moreover, there is little gain in efficiency from using the PDL

model, but it does hele to identify the appropriate lag length for each technology. Thus, except

for the discussion of lag length, I focus on the unrestricted model in the discussion in the text.

For reference, complete PDL results are presented here. Supplementary Table 13 shows the

results for government R&D for both individual years and the cumulative effect, and

Supplementary Table 14 shows the cumulative effects of the various controls included in each

model.

The main differences between the unrestricted results and the PDL results are as follows.

First, for biofuels, the FD results suggest a cyclical pattern, with strong effects also found in

years two and four in the first differenced model before picking up again in year t-9. In contrast,

the PDL model suggests a more gradual effect of energy R&D, with the effect initially peaking




44

in year t-2, and gradually fading until recurring in years nine and ten. However, the long-run

cumulative effects are similar in both models.

Supplementary Note 6. Detailed Results and Methodology for Patent Citations

Linking publications to eventual patents citing these publications is important, as creating

a new usable technology is the ultimate goal of any research funding. Here we provide detailed

results from the patent citation hazard regression, as well as a complete description of the

methodology used to simulate the effect of an additional one million dollars of public R&D

spending in Figure 3 of the main text.

Supplementary Table 15 presents the results of the patent hazard regression (equation 5

in the methods section), showing that the coefficients on citation lags are significant at the one

percent level, except for the squared term on wind energy. To interpret the time to citation, panel

A of Supplementary Figure 6 illustrates how the annual probability of citation changes over time.

For the post-2000 sample (shown using solid lines), the annual probability peaks between 8-10

years after publication. The peak shifts out to 11-14 years when considering the full sample

(dashed lines). As both the average and median lag between initial application and grant for

patents in our sample is 5 years, this means that patents citing these articles are filed 3-5 after

publication of the article in the post-2000 sample. Panel B shows the cumulative probability of

citation, which begins to grow rapidly 4-6 years after publication, and levels out about 12-14

years afterwards in the post-2000 sample.

The increased probability of an NPL citation resulting from an additional one million

dollars of new energy R&D funding depends on both the number of articles induced by this




45

R&D each year and on the probability of an article from any given year being cited in the future.

Thus, Figure 3 in the main text use the results of both our R&D estimation and the NPL citation

regression (IV results in Supplementary Table 9 and Supplementary Table 15):

(1) 𝑄𝑄𝑄𝑄𝑖𝑖𝑖𝑖,𝑡𝑡𝑡𝑡 = ∑ 𝛽𝛽𝛽𝛽𝑡𝑡𝑡𝑡−𝑠𝑠𝑠𝑠𝑅𝑅𝑅𝑅𝑖𝑖𝑖𝑖,𝑡𝑡𝑡𝑡−𝑠𝑠𝑠𝑠𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠=0 + ∑ 𝛄𝛄𝛄𝛄𝐭𝐭𝐭𝐭−𝐬𝐬𝐬𝐬𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝑖𝑖𝑖𝑖,𝑡𝑡𝑡𝑡−𝑠𝑠𝑠𝑠𝑇𝑇𝑇𝑇

𝑠𝑠𝑠𝑠=0 + ∑ 𝛅𝛅𝛅𝛅𝐭𝐭𝐭𝐭−𝐬𝐬𝐬𝐬𝐗𝐗𝐗𝐗𝑖𝑖𝑖𝑖,𝑡𝑡𝑡𝑡−𝑠𝑠𝑠𝑠𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠=0 + 𝛼𝛼𝛼𝛼𝑖𝑖𝑖𝑖 + 𝜂𝜂𝜂𝜂𝑡𝑡𝑡𝑡 + 𝜖𝜖𝜖𝜖𝑖𝑖𝑖𝑖,𝑡𝑡𝑡𝑡

(2) ℎ(𝑡𝑡𝑡𝑡) = exp(𝛼𝛼𝛼𝛼0 + 𝛼𝛼𝛼𝛼1𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 + 𝛼𝛼𝛼𝛼2𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐2 +𝛼𝛼𝛼𝛼3𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑚𝑚𝑚𝑚𝑐𝑐𝑐𝑐𝑡𝑡𝑡𝑡𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 + 𝛄𝛄𝛄𝛄𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐏𝐢𝐢𝐢𝐢,𝐭𝐭𝐭𝐭)

Let 𝛃𝛃𝛃𝛃 = [𝛽𝛽𝛽𝛽𝑡𝑡𝑡𝑡 𝛽𝛽𝛽𝛽𝑡𝑡𝑡𝑡−1 ⋯ 𝛽𝛽𝛽𝛽𝑡𝑡𝑡𝑡−𝑇𝑇𝑇𝑇] represent a row vector of coefficients on contemporary and

lagged R&D from Supplementary Equation (1). Using the results of Supplementary Equation

(2), the annual probability that a publication from year t receives a citation in year t + s can be

written as:

𝑤𝑤𝑤𝑤𝑠𝑠𝑠𝑠 = exp {𝛼𝛼𝛼𝛼0 + 𝛼𝛼𝛼𝛼1𝑠𝑠𝑠𝑠 + 𝛼𝛼𝛼𝛼2𝑠𝑠𝑠𝑠2}

Next, define a matrix WAnnual representing the annual probability that an article published

in year t, represented by the rows of the matrix, will be cited in year t + s, represented by the

columns of the matrix:

𝐖𝐖𝐖𝐖𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐀𝐚𝐚𝐚𝐚𝐚𝐚𝐚𝐚 =

⎣⎢⎢⎢⎡𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤1 𝑤𝑤𝑤𝑤2

𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤1 𝑤𝑤𝑤𝑤0

⋯ 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀−1

⋱ 𝑤𝑤𝑤𝑤0 ⎦

⎥⎥⎥⎤

Similarly, the cumulative probability of a publication in year t receiving a citation by year t + s is

given by the following matrix:

𝐖𝐖𝐖𝐖𝐏𝐏𝐏𝐏𝐀𝐀𝐀𝐀𝐂𝐂𝐂𝐂𝐀𝐀𝐀𝐀𝐚𝐚𝐚𝐚𝐚𝐚𝐚𝐚𝐭𝐭𝐭𝐭𝐢𝐢𝐢𝐢𝐂𝐂𝐂𝐂𝐂𝐂𝐂𝐂 =

⎣⎢⎢⎢⎡𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤0 + 𝑤𝑤𝑤𝑤1 𝑤𝑤𝑤𝑤2 + 𝑤𝑤𝑤𝑤1 + 𝑤𝑤𝑤𝑤0

𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤1 + 𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤0

⋯ 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀 + 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀−1 + … + 𝑤𝑤𝑤𝑤0 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀−1 + 𝑤𝑤𝑤𝑤𝑀𝑀𝑀𝑀−2 + … + 𝑤𝑤𝑤𝑤0

⋱ 𝑤𝑤𝑤𝑤0 ⎦

⎥⎥⎥⎤

Using these matrices, the product βWAnnual yields a row matrix with the increase in the

annual probability of an NPL citation each year after an additional one million dollars of energy




46

R&D, and the product βWCumulative yields a row matrix with the increase in the cumulative

probability resulting from an additional one million dollars energy R&D.

Supplementary Note 7. Additional Results on Diminishing Returns to R&D

Table 2 in the main paper provides results of regressions designed to test for the

possibility of diminishing returns to government R&D spending. To provide further

interpretation of the results including squared R&D, here we show how the effect of R&D varies

as the level of R&D spending varies. When adding a quadratic term, the marginal effect of R&D

will be a function of the R&D spending in a given year. Supplementary Figure 7 evaluates the

marginal effect of energy R&D for a given year for various lags, and Supplementary Figure 8

shows the cumulative effect. In each, these effects are evaluated for average levels of energy

R&D spending, as well as for the 25th, 75th, and 90th percentile. If diminishing returns are a

concern, we should expect to see lower marginal effects in the upper percentiles.

Our results provide almost no support for diminishing returns. In Supplementary Figure

7, we see slightly higher annual marginal effects for the 90th quantile of solar energy R&D for

most years, and for biofuels and energy efficiency in the later years. Only for wind do we see

the annual marginal effects fall for the highest quantiles, and only in the middle years.

Supplementary Figure 8 provides more insight, showing how the cumulative effect of

energy R&D varies over time for various levels of R&D funding. For biofuels and energy

efficiency, the cumulative effect is nearly identical across quantiles until the later years. This

would be consistent with the notion that large increases in R&D result in positive spillovers

providing dividends in future years, such as by attracting more researchers into the field. For




47

energy efficiency R&D, the cumulative effect for the 90th percentile is below those of other

percentiles through year three before catching up, raising the possibility of a period of

adjustment for large increases in energy R&D.

Solar energy provides evidence of positive spillovers from large energy increases, with a

significant positive quadratic terms. As shown in Supplementary Figure 8, the cumulative effect

of R&D is lowest for the 25th percentile, increasing significantly for each higher percentile.

Finally, the only technology showing any evidence of long run diminishing returns is wind.

Here, we see that the cumulative effect of wind energy is slightly larger in the 90th percentile in

early years, but soon becomes lower than the cumulative effect of other percentiles.

Finally, our second test of diminishing returns asks whether R&D spending is less

effective after large increases, represented by at least a doubling of energy R&D budgets for a

given technology. Supplementary Figure 9 shows that the number of countries choosing to

double energy R&D in a given year is generally equally dispersed across time, with one or two

countries per year experiencing such increases. The one exception is the large number of

countries choosing to double energy R&D as part of stimulus packages in 2009 and 2010.

Overall, between 6 and 12 percent of all country/year observations from 1981-2011 include a

doubling of energy R&D.




48

Supplementary References

1. Beck, T., Clarke, G., Groff, A., Keefer, P., & Walsh, P. New tools in comparative

political economy: The Database of Political Institutions. World Bank Economic Review

15, 165-176 (2001).

2. OECD-EPAU. Renewable Energy Policy Dataset, version February 2013. Compiled by

the OECD Environment Directorate’s Empirical Policy Analysis Unit (Johnstone, N.,

Haščič, I., Cárdenas Rodríguez M., Duclert, T.) in collaboration with an ad hoc research

consortium (Arnaud de la Tour, Gireesh Shrimali, Morgan Hervé-Mignucci, Thilo Grau,

Emerson Reiter, Wenjuan Dong, Inês Azevedo, Nathaniel Horner, Joëlle Noailly, Roger

Smeets, Kiran Sahdev, Sven Witthöft, Yunyeong Yang, Timon Dubbeling) (2013).

3. Johnstone, N., Hascic, I & Popp, D. Renewable energy policies and technological

innovation: Evidence based on patent counts. Environmental and Resource Economics

45(1), 133-155 (2010).

4. Baccini, L. & Urpelainen, J. Legislative fractionalization and partisan shifts to the left

increase the volatility of public energy R&D expenditures. Energy Policy 46, 49-57

(2012).

5. Crespi, G.A. & Geuna, A. An empirical study of scientific production: A cross country

analysis, 1981-2002. Research Policy 37, 565-579 (2008).

6. International Energy Agency. Renewable Energy: Market & Policy Trends in IEA

Countries (OECD/IEA 2004).

7. Cheon, A. & Urpelainen, J. Oil prices and energy technology innovation: An empirical

analysis. Global Environmental Change 22, 407-417 (2012).




49

8. Verdolini, E. & Galeotti, M. At home and abroad: An empirical analysis of innovation

and diffusion in energy technologies. Journal of Environmental Economics and

Management 61, 119-134 (2011).

9. Dechezleprêtre, A., & Glachant, M. Does foreign environmental policy influence

domestic innovation? Evidence from the wind industry. Environmental and Resource

Economics 58, 391-413 (2014).

10. Nesta, L., Vona, F. & Nicolli, F. Environmental policies, competition, and innovation in

renewable energy. Journal of Environmental Economics and Management 67, 396-411

(2014).

11. Dechezleprêtre, A. & Popp, D. Fiscal and regulatory instruments for clean technology

development in the European Union. CESifo Working Paper # 5361 (2015).


economic analysis of scientific publications and …...supplementary figure6 – probability of...

Documents