City, University of London Institutional Repository

Citation: Banal-Estanol, A., Jofre-Bonet, M. and Lawson, C. (2015). The Double-Edged Sword of Industry Collaboration: Evidence from Engineering Academics in the UK. Research Policy, 44(6), pp. 1160-1175. doi: 10.1016/j.respol.2015.02.006

This is the accepted version of the paper.

This version of the publication may differ from the final published version.

Permanent repository link: https://openaccess.city.ac.uk/id/eprint/13878/

Link to published version: http://dx.doi.org/10.1016/j.respol.2015.02.006

Copyright: City Research Online aims to make research outputs of City, University of London available to a wider audience. Copyright and Moral Rights remain with the author(s) and/or copyright holders. URLs from City Research Online may be freely distributed and linked to.

Reuse: Copies of full items can be used for personal research or study, educational, or not-for-profit purposes without prior permission or charge. Provided that the authors, title and full bibliographic details are credited, a hyperlink and/or URL is given for the original metadata page and the content is not changed in any way.

City Research Online: http://openaccess.city.ac.uk/ publications@city.ac.uk

City Research Online

1

TheDouble‐EdgedSwordofIndustryCollaboration:

EvidencefromEngineeringAcademicsintheUK

AlbertBanal‐Estañol,ab*MireiaJofre‐BonetaandCorneliaLawsoncd

aCityUniversityLondon,DepartmentofEconomics,NorthamptonSquare,London,EC1V0HB,UKbUniversitatPompeuFabra,DepartmentofEconomicsandBusiness,RamonTriasFargas25‐27,

08005Barcelona,SpaincBRICK,CollegioCarloAlberto,ViaRealCollegio30,10024Moncalieri Turin ,Italy

dUniversityofNottingham,SchoolofSociologyandSocialPolicy,UniversityPark,Nottingham,NG72RD,UK

*Correspondingauthor:AlbertBanal‐Estañol,UniversitatPompeuFabra,DepartmentofEconomicsandBusiness,RamonTriasFargas25‐27,08005Barcelona,Spain,email:albert.banalestanol@upf.edu,Tel.: 34935422871,Fax: 34935421746

Emailaddresses:albert.banalestanol@upf.edu(Banal‐Estañol),mireia.jofre‐bonet@city.ac.uk(Jofre‐Bonet),cornelia.lawson@nottingham.ac.uk(Lawson).

Abstract

Thispaperstudiestheimpactofuniversity‐industrycollaborationon academic research output. We analyze the channels throughwhichthedegreeofindustrycollaborationmaybeaffectingresearchoutput. We exploit a unique longitudinal dataset on all theresearchers in all the engineering departments of 40 majoruniversities in the UK for the last 20 years. We use an innovativemeasure of collaboration based on the fraction of public researchgrants that include industry partners. Our empirical findingscorroborate that therelationshipbetweencollaborationdegreeandpublicationratesiscurvilinear,andshedsomelightontheselectionmechanismsatwork.Ourresultsarerobust toseveraleconometricmethods,measuresofresearchoutput,andsubsamplesofacademics.

Keywords: Industry‐science links, research collaboration, basic vs. appliedresearch.JELcodes:O3,L31,I23

2

1 Introduction

Inamoderneconomytransformingscientificresearchintocompetitiveadvantagesis

essential. In theUS,extensivecollaborationbetweenuniversitiesand industry, and the

ensuingtransferofscientificknowledge,isviewedasoneofthemaincontributorstothe

successfultechnologicalinnovationandeconomicgrowthofthepastthreedecades Hall,

2004 .Atthesametime,accordingtoaEuropeanCommissionreport 1995 ,insufficient

interactionbetweenuniversitiesandfirmsintheEUhasbeenoneofthemainfactorsfor

theEU’spoorcommercialandtechnologicalperformanceinhigh‐techsectors.Nowadays,

increasinguniversity‐industry collaboration is a primarypolicy aim inmostdeveloped

economies.1

Theincreasedincentives or,assomesay,pressure tocollaboratewithindustrymay

have controversial side effects on the production of scientific research itself. Nelson

2004 , amongmany others, argues that the existence of industry involvement might

delay or suppress scientific publication and the dissemination of preliminary results,

endangeringthe“intellectualcommons”andthepracticeof“openscience” Dasguptaand

David,1994 .FloridaandCohen 1999 arguethatindustrycollaborationmightcomeat

theexpenseofbasicresearch:growingtieswithindustrymightbeaffectingthechoiceof

researchprojects,“skewing”academicresearchfromabasictowardanappliedapproach.

Academicsthatcontributetoknowledgeandtechnologytransfer,ontheotherhand,

maintain that theexistenceof industry collaborationcomplements theirownacademic

researchbysecuring funds forgraduatestudentsand labequipment, andbyproviding

themwithideasfortheirownresearch Lee,2000,AgrawalandHenderson,2002 .Siegel

etal. 2003 , forexample,reportthat“ s omescientistsexplicitlymentionedthatthese

interactions improved the quantity and quality of their basic research.” Ideas sourced

from industry may thus expand traditional research agendas Rosenberg, 1998 ,

benefittingtheoverallscientificperformanceofresearchers.

These opposing claims raise a long‐standing question for academic research: Does

collaborationwithindustryincreaseordecreasepublicationrates?Previousresearchon

this issuehasmostlyusedpatentingasameasureof industrycollaboration seeGeuna

and Nesta, 2006, and Baldini, 2008, for reviews . The evidence is somewhat mixed,

rangingfromthenegativeeffectsofpatentingonresearchoutputreportedinsurveysof

1 Inthe1980s,theUSintroducedaseriesofstructuralchangesintheintellectualpropertyregimeaccompaniedbyseveral incentive programs, designed specifically to promote collaboration between universities and industryLee,2000;Moweryetal.,2001 .Almost30yearson,manyelementsoftheUSsystemofknowledgetransferhavebeenemulatedinmanyotherpartsoftheworld seee.g.,theUKGovernment’sWhitePaper“TheFutureofHigherEducation,”2003 .

3

academic scientists Blumenthal et al., 1996 , to no effect in some of the econometric

studies Agrawal andHenderson, 2002 to even a positive relationship in some of the

recentevidence Azoulayet al., 2009;Breschi et al., 2008;FabrizioandDiMinin,2008;

Stephanetal.,2007;vanLooyetal.,2006 .

Thispaperarguesthatacademicresearchoutputisaffectednotonlybytheexistence

of linkswith the industry but also by the degree of industry collaboration, i.e., by the

proportion of or the share of time spent on projectswith industry involvement.2 By

exploring the channels through which industry collaboration affects publications, our

conceptual frameworksuggests that therelationshipbetweencollaborationdegreeand

publicationratesisneitherincreasingnordecreasingbutiscurvilinear,describedbyan

invertedU‐shapedcurve.Asaresult,researchoutputshallbemaximizedatintermediate

degrees of collaboration, i.e., when the industry is involved in some but not in all the

projectsoftheacademic.

Our empirical analysis uses an innovative measure of degree of industry

collaboration based on the fraction of publicly funded research projectswhich include

industrypartners.3 Incontrasttopatents,thismeasureiscontinuousinnature,andsois

abletoproxynotonlyfortheexistencebutalsoforthedegreeofindustrycollaboration.

In addition, collaborative links through joint research, consulting or training

arrangements aremorewidespread D’Este and Patel, 2007 and aremore important

knowledge transfer channels than patents, licenses, and spin‐offs, according to both

academics Agrawal and Henderson, 2002 and firms Cohen et al., 2002 . Data on

research collaborations also provide a more continued assessment of the level of

interactionwith industrythanmeasurementsbasedonthenumberofpatents.Possibly

duetothelackofcomparabledata,theliteraturehaspaidlittleattentiontothesemore

collaborativeformsofuniversity‐industryinteraction.

Ourmeasure of collaboration is constructed exploiting comprehensive information

from themainUK government agency for funding in engineering, theEngineering and

PhysicalSciencesResearchCouncil EPSRC ,whichdistinguishesbetweencollaborative

andnon‐collaborativeresearchgrantsbasedontheinvolvementofindustrypartners.In

addition to research funds, we compiled a unique, longitudinal dataset containing

2 Ournotionof“degreeof industry collaboration” is inspiredbythenotionof“degreeof research collaboration”used in bibliometric studies. As shown by Subramanyam 1983 , the degree of research collaboration is usuallydefinedasthenumberofmulticoauthoredpapersoutofthetotalnumberofpapers singleandmulticoauthored .

3 The presence of industry partners in public research grants might not be a perfect proxy for the degree ofcollaborationwithindustry,asthereareotherchannelsof interaction.Theinclusionofprivatefirmsaspartnersinthesegrants,however,ishighlycorrelatedwithobtainingdirectfundingfromindustry Meissner,2011 .

4

research output publications , patents, and other individual characteristics for all

academics employed in all the engineering departments of 40 major UK universities

between1986and2007.Sinceourdatasetcontainsthemajorityofacademicengineers

intheUK,ourresultsarenotdrivenbythemostsuccessfuloracademicinventorsalone.

Infact,wecantestwhethertheeffectsdifferacrossobservedcategoriesofresearchers.

Still,theobserveddegreesofcollaborationarenotexogenouslydetermined,butare

theresultofindividualandmutualchoicesinatwo‐sidedmarketofacademicsandfirms

(Mindruta, 2013; Banal‐Estañol et al, 2014). Unobserved characteristics of the

researchers may affect not only their degree of collaboration but also their academic

productivity,therebyinfluencingtheshapeofthecollaboration‐publicationrelationship.

Ourconceptualframeworkanalyzesthepotentialselectionmechanismsatwork,andthe

direction of the biases one might incur if these mechanisms were ignored in the

estimation. As empirical strategy, to deal with the endogeneity problem, we use fixed

effects and instrumental variable techniques and, to take into account the dynamic

natureofthepublications e.g.,Aroraetal.,1998;AgrawalandHenderson,2002 ,weuse

a dynamic panel data approach. By comparing instrumented and non‐instrumented

regressionsweshallalsoshedmorelightontheselectionmechansimsinplace.

Thepaperisorganizedasfollows.Insection2weprovidetheconceptualframework.

Insection3wedescribethedatasetandinsection4weintroduceourempiricalstrategy.

Section5presentsourresults.Section6discussesandconcludes.

2 Conceptualframework

Our conceptual framework is based on the idea that the degree of industry

collaboration of an academic affects the main determinants of her scientific output,

namely, i the quality and quantity of her ideas; ii the time and attention she can

devotetodevelopingtheseideasandtransformingthemintopapers; iii theamountof

resourcesshehasavailable;4 and iv theexistenceofconstraints onthescopeand/orin

thedisseminationofresearchresults Stephan,1996,2012 .Inthefirstsubsectionbelow,

we discuss the channels through which the degree of industry collaboration may be

affectingresearchoutput.Inthesecondsubsection,weconsiderthecharacteristicsofthe

academics thatmaybeaffectingboth theirobserveddegreesofcollaborationandtheir

4 The availability of research funding is important for scientists in all academic disciplines, but especially in resource-intensive fields such as engineering (Stephan, 1996, 2012). Several recent studies have documented a positive impact of public grants on research performance (Jacob and Lefgren, 2011; Benavente et al., 2012).

5

researchoutput.Wedescribethepotentialselectionmechanismsatwork,allowingusto

identifypotentialbiasesintheestimation.

2.1 Effectsofthedegreeofindustrycollaborationonresearchoutput

Collaborationwithindustrycanboostresearchoutputforatleasttworeasons.First,

collaborationcanexpandacademics’researchagendasandimprovethepoolofresearch

ideas Rosenberg,1998 .Mansfield 1995 showsthatasubstantialnumberofpublicly

sponsored researchprojects stem from industrial problems encountered in consulting.

Collaboration helps academics gain new insights for their own research and test the

practical application of their theoretical ideas Lee, 2000 . The generation and/or

refinement of ideas through puzzle‐solving may in turn improve research outcomes

becausetheresultingideascanbetransformedintomoreand/orbetteracademicpapers.

Second, industry collaboration can expand the availability of financial resources.

According to survey evidence in Lee 2000 , two of the most important reasons for

academics to collaborateare to secure funds forgraduate studentsand labequipment,

andtosupplementfundsfortheiracademicresearch.Inrecentyears,industryhasbeen

identified as an evenmore important sourceof funding for academic research. Private

financial support is important in light of progressive declines in direct government

funding OECD,2010 andofmorecompetitiveresearchenvironments Stephan,2012 .

Nevertheless, an increase in the degree of industry collaboration is not necessarily

positive for research output for at least five reasons. First, although collaborationmay

enhancethepoolofresearchideas,theremightbedecreasingreturnstoscaleassociated

tothegenerationoftheseideas.HottenrottandLawson 2014 showthatresearchunits

that receive larger shares of funding originating from industry are alsomore likely to

develop ideas stemming from private partners, suggesting that the pool of ideas for

higher degrees of collaboration is indeed larger. But, at high degrees of collaboration,

ideasatthemarginmaynotbeofthesamequality,ormaynotevenbeworthpursuing,

andmaythusnotresultinthesameincreaseinpublicationratesastheideasobtainedat

lowdegreesofcollaboration.

Second, a high number of ideas, conceivably available from higher degrees of

collaboration,mayalsocreateattentionproblems.Accordingtoattention‐basedtheories

ofthefirm,decision‐makersinanyorganizationneedto“concentratetheirenergy,effort

andmindfulness on a limitednumber of issues” Ocasio, 1997 .As arguedby Laursen

andSalter 2006 inthecontextof innovativefirms,iftherearetoomanyideas,fewof

6

themreceive therequired levelof timeoreffort tobedevelopedseriously.Asa result,

highlevelsofcollaborationmaygeneratemanyideasbutfewacademicpapers.

Third,collaborativeprojectsusuallyrequiretimeforcoordination,organization,and

interaction.Collaborationwithaprivatepartnermayalsocomewith“stringsattached”in

the form of academic consulting or commercial activities. The general duties of the

academics,andresearchinparticular,mightbecompromisedbyanincreaseinthetime

allocated to development, consulting or commercialization Florida and Cohen, 1999 ,

thusreducingscientificpublication.

Fourth, collaborationmay affect the selection of topics andmethodologies Florida

and Cohen, 1999 . As argued in Trajtenberg et al. 1997 , industry research and

development tends to be directed at commercial success, while university research

generally focuses on solving fundamental scientific questions. Thus, research that

appeals to industry partners may not necessarily be close to the research frontier

Rosenberg and Nelson, 1994 , and may be less likely to result in top academic

publications.Thisisespeciallythecaseforacademicswithhighdegreesofcollaboration

who may get locked in service provision for industry Meyer‐Krahmer and Schmoch,

1998 .

Finally, firms’ commercial interests might impose constraints on the publication

activityof collaboratingacademics, especially those that collaborateextensively.Firms’

commercial interests may push firms to include non‐disclosure clauses that delay or

suppress scientific publication Nelson, 2004 . Czarnitzki et al. 2014 indeed find

empiricalevidence that thepercentageof researchers thatcomplainaboutsecrecyand

publicationdelayislargerforresearcherssponsoredbyindustry.

As summarized in the upper part of Figure 1, collaboration can have positive and

negative effects on the factors driving academic research. The relative magnitudes of

these effects and their ultimate impact on publications change with the degree of

collaboration.Fromourdiscussion,weexpectthenegativeeffectstoberelativelymore

important and, thus, to dominate for high degrees of collaboration while the positive

ones shall dominate for low degrees of collaboration. We therefore anticipate the

relationship between collaboration degree and publication rates to have an inverted

U‐shapeandresearchoutputtobemaximizedat intermediatedegreesofcollaboration,

i.e.,whenindustryisinvolvedinsomebutnotalltheprojectsoftheacademic.

7

2.2 Mechanismsinfluencingthecollaboration‐researchoutputrelationship

This subsection discusses other mechanisms that may affect the relationship

betweenindustrycollaborationandresearchoutputdescribedintheprevioussubsection.

Wearguethatthedegreesofcollaborationarenotexogenouslydetermined,butarethe

result of individual andmutual choices in a two‐sided ‘market’ of academicsand firms

(Mindruta,2013;Banal‐Estañoletal,2014).5 Observedandunobservedcharacteristics

oftheresearchers(suchasseniority,abilityorskills)mayaffectnotonlytheirdegreeof

collaborationbutalsotheiracademicproductivity, thereby influencingtheshapeof the

collaboration‐publication relationship. In this subsection, we focus on the unobserved

characteristics anddescribe thepotentialmechanismsatwork, by ‘type’ of researcher,

andthedirectionofthebiasesonemightincurifthesemechanismswereignoredwhen

estimatingthecausaleffectofdegreeofcollaborationonresearchoutput.

Notice first that there are inherent characteristics of the researchers that make

thembothmore likely topublish andmore likely to finda goodpartner. For example,

ability or talent is important for research but it is also highly valued by firms when

searchingforresearchpartners(seee.g.,Blumenthaletal.,1996).Asaresult,highlyable

academicsmayenduphavinghigherdegreesofcollaborationandbetterpartnersthan

less‐able academics. As argued by Blumenthal et al. (1986), “the most obvious

explanationfor[the]observedpositiverelation[betweencollaborationandpublication]

is that companies selectively support talented and energetic facultywhowere already

highlyproductive.”Throughtheir involvementwithmoreattractivepartners, theymay

endupproducingevenmoreacademicpapers.AsshowninBanal‐Estañoletal. 2013 ,

collaborativeprojectsgeneratemorepublicationsthannon‐collaborativeonesifandonly

if the industrial partners are highly productive. As a result, if one does not take into

account the presence of researcher “fixed effects” such as ability, the estimation may

generatepositivebiasesontheeffectofhighdegreesofcollaborationonpublications.

Time‐invariant individualcharacteristicsmayalso interactwith time‐variantones

and generate other biases. First, some of the talentedacademicsmay develop stronger

preferences for industry collaboration and better networking skills, thereby enjoying

higher‐quality interactionsandmoreknowledgeof theprivate sector.Weexpect these

5 Firmsalsoweighthebenefitsandcostsofcollaboratingwithacademicpartners(Hendersonetal.,1998;SalterandMartin,2001;Cohenetal.,2002;LinkandScott,2005,Laursenetal.,2011).Firmsreporttocollaboratetogetaccesstonewuniversityresearchanddiscoveries(Lee,2000),butarealsoconcernedwiththeorganizationalandinstitutionalstructure,andtheexistenceoftheopenscienceculture,inacademia DasguptaandDavid,1994 .Anindividual firm’sdecisiontocollaboratedependsonitsabsorptivecapacity VeugelersandCassiman,2005 , itssize,andwhetheritadoptsanopensearchstrategy MohnenandHoareau,2003;LaursenandSalter,2004 .

8

“industry‐savvy” researchers to have higher degrees of collaboration and, at the same

time,tochooseandtobechosenbybetterindustrialpartnersbecausetheyscreenbetter

andtheyaremoreappealingacademicpartners.Thankstotheirinvolvementwithmore

attractive industry partners, these academics are again predicted to produce more

academic papers than those at low degrees of collaboration. In sum, more engaged

researchersmayhavebothhigherdegreesof collaborationand–because theyendup

being matched with better partners – more research output than researchers lacking

collaboration taste and networking skills.Wemay thus observe a positive bias at the

high‐endof thedegreeofcollaboration(thosemore likelypopulatedby industry‐savvy

researchers) and, symmetrically, a negative bias at low degrees of collaboration

(populatedbynon‐industry‐savvyresearchers).

Second, some of the talented academics may becomemore output‐driven in their

collaborationchoices.Weexpecttheseresearcherstopublishmoreand,atthesametime,

prefer intermediate degrees of collaboration, as they are more likely to identify and

ponder the trade‐offs described in the previous subsection (e.g., new ideas vs. time

constraints). That is, these highly able, output‐driven researchers may end up having

mid‐rangedegrees of collaboration,while less able, less output‐driven academics,may

end up having insufficient or excessive degrees of collaboration that negatively affect

theiracademicperformance.Inaddition,thehighlyablewillalsoenduppartneringwith

moreattractivepartners,producingevenmorepapers.Wemaythusobserveapositive

biasonthemid‐rangedegreesofcollaborationandanegativebiasinboththelow‐and

high‐endsofthedegreeofcollaboration.

Third,someofthetalentedresearchersmayalsobecomemoreselectiveovertime,

given the increased pool of potential industry projects they have available. While

collaborationwithhighlyablescientistsisthemostbeneficial forfirms,companiesfind

academic partners across awhole quality‐range of researchers and departments,with

the majority of industry funding going to universities of medium research quality

(Mansfield, 1995). Goldfarb (2008) argues that star‐scientists have many more

opportunities for funding but it is the less‐able researchers who typically engage in

programssponsoredbymission‐orientedagentssuchasfirms.Less‐ableacademicsmay

have to accept any industry support in order to maintain or increase their funding

(Carayol,2003).Insum,talented,successfulresearchersmaybeinapositiontobemore

selective andengageonly in collaborationswithhigh‐quality industrypartners.Due to

theexistenceofthesetypesofacademics,wemayobserveapositivebiasatthelow‐end

9

of the degree of collaboration and a corresponding negative bias in the mid‐ and

high‐ends.

The selectionmechanismsdescribed in this section are summarized in the lower

part of Figure 1. As explained in more detail in the empirical strategy section, these

mechanismshavefundamentalimplicationsforanyattempttoestimatetherelationship

between the degree of industry collaboration and research output. The degree of

collaborationis“endogenous,”anditisaffectedbyobservedandunobservedindividual

characteristicsthatalsoaffectresearchoutput.Ourempiricalapproachshallcontrolfor

time‐invariantandtime‐variantobservedcharacteristics.Weshallalsotakeintoaccount

theexistenceofunobservedcharacteristicsinfluencingtheselectionmechanismsatwork.

Notdoingsowouldresultinabiasedestimateoftheimpactofdegreeofcollaborationon

research output. In our empirical strategy section, we explain how we address these

endogeneityconcerns.

3 Data

Inthissection,weprovideadetailedaccountofhowwecreatedourdataset.Forthis

study, we built a unique longitudinal dataset containing individual characteristics,

publications, research funds, and patents for all researchers employed in all the

engineeringdepartmentsof40majorUKuniversitiesbetween1986and2007 seeTable

1 fora listof theuniversities .Through theBritishLibrary,wesearched foruniversity

calendars andprospectusesprovidingdetailed staff information for all theuniversities

withengineeringdepartmentsintheUK.6 Ourfinalsamplecontainsalltheuniversities

thathadcalendar informationavailable, includingall theuniversitiesthataremembers

oftheprestigiousRussellGroup,acoalitionof24research‐intensiveUKuniversities,as

wellas16othercomprehensiveortechnicaluniversities.7

Weretrievedtheacademics’namesandranksforalltheyearsfrom1986to2007.We

focusedonacademicstaffcarryingoutbothteachingandresearchanddidnotconsider

research officers or teaching assistants. We followed the researchers’ career paths

6 ByActofParliament,theBritishLibraryisentitledtoreceiveafreecopyofeveryitempublishedintheUK.ThesedataweresupplementedwithinformationfromtheInternetArchive,anot‐for‐profitorganizationthatmaintainsafreeInternetlibrarycommittedtoofferingaccesstodigitalcollections.Theircollectiondatesbackto1996andenabledustoretrieveinformationfromoutdatedInternetsites.

7 Weidentifiedtheinitialsetofengineeringdepartmentsfromthe1996and2001ResearchAssessmentExercisesRAEs .We did not find staff information for eight institutionswhich are similar to the 16 non‐Russell groupuniversities inoursample.Wedidnotconsideranyofthe39post‐92universitieseither,asthesewerenot fullresearch institutions for all the years considered in our analysis. We also excluded the Open University andCranfieldUniversitywhich,asdistanceandpostgraduateinstitutions,respectively,haveaverydifferentstructure.

10

betweenthedifferentuniversitiesbymatchingnamesandsubjectareasandbychecking

the websites of the researchers. Academics leave and join or rejoin our dataset at

different stages in their career, when they move to or from abroad, industry,

departmentsotherthanengineering e.g.,chemistry,physics ,oruniversitiesthatarenot

part of ourdataset, resulting in anunbalancedpanel. They represent thebasis for our

datacollectionandenableustoretrieveinformationonpublications,researchfunds,and

patents.

Ourfinalsamplecontainsinformationon3,991individuals.Thefinalsampleexcludes

all inactive researchers those with neither publications nor funds during the entire

sampleperiod andresearcherswhowerepresentforlessthansixconsecutiveyearsso

thatallofour stock variablescouldbecreated.8 Wedescribebelowoursourcesand

measuresofresearchoutput ourdependentvariable ,degreeofindustrycollaboration

our main independent variable , as well as funding, patents, and other individual

characterizingvariables.WeprovidesummarystatisticsinthefirstpanelofTable2.

Research output. Data on publications were obtained from the ISI Science Citation

Index SCI . The number of publications in peer‐reviewed journals is not the only

measurebutisthebestrecordedandthemostacceptedmeasureforresearchoutputas

they are essential for gaining scientific reputation and for career advancements

Dasgupta andDavid, 1994 .We collected information on all the articles published by

researchers inourdatabasewhile theywereemployedatoneof the institutions inour

sample.MostentriesintheSCIdatabaseincludedetailedaddressdatathatallowedusto

identify institutional affiliations and unequivocally assign articles to individual

researchers.9

Asamainmeasureof researchoutput foreachresearcher ineachyear,weuse the

normal count of publications countit , i.e., the number of publications in t on which

researcheriisnamedasanauthor.Publicationcounts,however,mightbemisleadingfor

articles with a large number of authors and may not reflect a researcher’s effective

productivity. Therefore, we also use the co‐author‐weighted count of publications

co‐authorweightedcountit ,whichweobtainbyweightingpublicationsbytheinverse

ofthepublication’snumberofco‐authors.

8 Estimationsconsideringashortertimewindowofjustthreeconsecutiveyearsareusedintherobustnesschecks.Thedescriptivestatisticsaswellastheempiricalresultsareverysimilartothoseofthemainestimation.

9 Publicationswithoutaddressdatahadtobeignored.However,weexpectthismissinginformationtoberandomandtonotaffectthedatasystematically.

11

We also separate the count of publications by type of academic research basic or

applied . To construct these measures, we use the Patent Board formerly CHI

classification version2005 ,developedbyNarinetal. 1976 andupdatedbyKimberley

HamiltonfortheNationalScienceFoundation NSF .Basedoncross‐citationmatrices,it

characterizes the general research orientation of journals, distinguishing between 1

appliedtechnology, 2 engineeringandtechnologicalscience, 3 appliedandtargeted

basicresearch,and 4 basicscientificresearch.Godin 1996 andvanLooyetal. 2006

reinterpretedthecategoriesas 1 appliedtechnology, 2 basictechnology, 3 applied

science,and 4 basicscience;andgroupedthefirsttwoas“technology”andthelasttwo

as“science.”Weusethenormalcountofpublicationsineachofthesetwocategoriesand

denote them“applied” appliedcountit and“basic” basiccountit , respectively.Due to

theappliedcharacteroftheengineeringsciences,74%ofallpublicationsareapplied.

Degreeof industry collaboration.Ourmeasureof industry collaboration isbasedon

grantsawardedbytheEngineeringandPhysicalSciencesResearchCouncil EPSRC ,the

mainUKgovernmentagencyforresearchinengineeringandthephysicalsciences,and

by far the largest provider of funding for research in engineering more than 50% of

overall third‐party funding . The EPSRC encourages but does not require academic

researcherstofindprivatepartnersfortheirresearchprojects.AsdefinedbytheEPSRC,

“CollaborativeResearchGrantsaregrantsledbyacademicresearchers,butinvolveother

partners.” Partners generally contribute either cash or “in‐kind” services to the full

economiccostoftheproject.10

Weobtained informationon all the grants awarded since1986. For each grant,we

collectedthestartyear,duration,totalamountoffunding,namesofprincipalinvestigator

PI andco‐investigators,grant‐receivinginstitution,andnamesofpartnerorganizations,

ifany. Inordertoconstructourproxyforthedegreeofcollaborationwith industrywe

use the presence of private partners. Our variable, which we name fraction of EPSRC

funding with industryit, represents the fraction of collaborative EPSRC funds of an

individualiinthefivepreviousyears i.e.,betweent‐4andt .Weuseafive‐yearwindow

toreflecttheprofileofanacademicintermsofherpaststreamoffunding.

Tobeprecise,thisvariablewasconstructedasfollows.Wedividedthetotalmonetary

incomeofeachgrantbetweenthePIandherco‐investigator s .Wetookintoaccountthe

10 TheEPSRCdoesnotfavorspecifictypesofacademicresearchoutput.Bothcollaborativeandnon‐collaborativegrantsareawardedbasedonpeer‐reviewandmonitoredthroughend‐of–awardreports.Thereare,however,nospecificmeasuresforevaluatingthesuccessoftheknowledgeexchangesbetweenscienceandindustry.

12

participationofall investigatorsbutpositivelydiscriminatedPIsbyassigningthemhalf

the grant value and splitting the remaining 50% among the co‐investigators. PIswere

assignedamajorpartastheyareexpectedtoleadtheproject.Weadditionallyspreadthe

grantvalueovertheawardperiod.11 Thiswasdoneinordertoaccountfortheongoing

benefitsandcostsoftheprojectandtomitigatetheeffectoffocusingallthefundsatthe

start of the project. Finally, for each year and for each researcher we computed the

fractionof“collaborative”fundsinthelastfiveyears,i.e.,thosethatincludedoneormore

industrypartners,overallEPSRCfundsreceivedinthesameperiod.

Funding. Asdiscussedintheconceptualframeworksection,theavailabilityoffunds

is an important factor for research output.We therefore created an indicator variable

hadsomeEPSRCfundingit thattakesvalue1ifacademicireceivedanyEPSRCfunding

collaborativeornot inthefivepreviousyearsand0otherwise.Weusedtheindicator

instead of the funding level because the latter is discipline‐specific some disciplines

requiremore funding than others . As for industry collaboration, we used a five‐year

windowtoreflecttheacademic’sprofile.

Patents. In the conceptual framework, we argued that the commercialization of

research resultsmight impose constraints on publication activity. Our analysis should

thereforecontrolforpatentactivity.Byincludingpatents,wealsoseparatetheeffectof

patentingfromtheeffectofcollaboratingwithindustry,asdefinedabove.Priorresearch

has considered patenting itself to be an indicator of a researcher’s involvement with

industry.Asa result, thebenefitsandcostsof collaborationmightalsoappear through

thepatentchannel.

We obtained patent data from the European Patent Office EPO database. We

collected those patents that identify the aforementioned researchers as inventors and

werefiledwhiletheywereemployedatoneofthe40institutions.Wenotonlyconsider

patentsfiledbytheuniversitiesthemselves,butalsothoseassignedtothirdparties,e.g.,

industryorgovernmentagents asshownbyLawson,2013b,52%ofacademicpatentsin

the UK are not owned by the university . The filing date of a patent was recorded as

representingtheclosestdatetoinvention.Sincethefilingprocesscantakeseveralyears,

11 Ifthegrantlastedtwoyears,wesplititequallyacrossthosetwoyears.Ifitlastedthreeormoreyears,thefirstand the last years which are assumed to not represent full calendar years received half the share of anintermediateyear.

13

wewereonlyabletoincludepatentspublishedby2007,hencefiledbefore2005.12 The

EPO only covers a subsample of patents filedwith the UK Intellectual Property Office

UKIPO . Nevertheless, the patents that are taken to the EPO are those with a higher

economicpotentialand/orquality MaursethandVerspagen,2002 andhavebeenused

inthepasttoanalyzeacademicpatentinginEurope seeLissonietal.,2008 .

Tomeasure the impactofpatentingon the timingofpublications,weuseadummy

variableindicatingwhethertheacademicifiledanypatentinthesameyear patentit ,or

in the two years preceding the publication patentit‐1 and patentit‐2 . Researchers in

Europe,unliketheUS,cannotbenefitfroma“graceperiod”andhencehavetowithhold

anypublication related to thepatentuntil thepatent application is filed.We therefore

expectalagofuptotwoyearsbetweeninventionandpublicationinajournal.

Individual characteristics. Research output might be linked to the researcher’s

personalattributessuchassex,age,education,andacademicrank.AcademicRankitisthe

only time‐variant observable characteristic in our dataset. Thus, we incorporate

information on the evolution of researchers’ academic status from lecturer to senior

lecturer,reader,andprofessor.LecturercorrespondstoanassistantprofessorintheUS,

whereasseniorlecturerandreaderwouldbeequivalenttoanassociateprofessor.

Wealsoinclude,asanadditionaltime‐variantcharacteristicofanacademicatagiven

point in time, her past publications. Indeed, as argued by Stephan 1996 , there is a

“cumulative advantage” in science that results in a dynamic relationship between past

andpresentpublicationoutput.

Interaction variables. The effect of degree of industry collaboration on research

output might differ across observed categories of academics. In order to investigate

whether the relationship differs between academics who have certain observed

characteristics, we create indicator variables that reflect i being above or below the

medianlifetimeshareofpublicationswithindustryco‐authors; ii beingaboveorbelow

themedianlifetimeshareofcollaborativegrants; iii beingaboveorbelowthemedian

amountoffundingduringthepreviousfiveyears; iv belongingtotheselectedRussell

group of universities; and, v being at an earlier stage of their careers as opposed to

beingseniorresearchers professors .Weallowresearcherstochangegroupswhenthey

12 Justlikepreviousstudies seee.g.,FabrizioandDiMinin,2008 ,dataconstructionrequiresamanualsearchinthe inventordatabase to identify entries thatwere truly the same inventor andexcludeotherswith similaroridenticalnames.Thiswasdonebycomparing theaddress, title, and technologyclass forallpatentspotentiallyattributabletoeachinventor.TheEPOdatabaseisproblematicinthatmanyinventionshavemultipleentries.Itwasthusnecessarytocompareprioritynumberstoensurethateachinventionisonlyincludedonceinourdata.

14

change universities or are promoted. Panels 2 to 6 in Table 2 present the descriptive

statisticsofthetwomainvariablesofinterestforthesesubsamples.

4 Empiricalstrategy

Thissectiondescribestheeconometricspecificationofourmodel,themethodsweuseto

estimateit,andtheinstrumentalvariablesweexploitinordertodoso.

4.1. Econometricspecification

According to our conceptual framework, the relationship between the degree of

collaborationandthepublicationsofanacademiccanbecurvilinear.Indeed,increasing

thedegreeofcollaborationwithindustrycanboostresearchoutput,ascollaborationmay

improvethepoolofresearchideasandexpandtheavailabilityoffinancialresources.But,

at high degrees of collaboration, it may also be negative, because there might be

decreasingreturnsassociatedtothegenerationofideas,extensiveindustryinvolvement

may impose timeconstraintsandcauseattentionproblems,andcommercialobjectives

and strategic behavior may push industrial partners to impose constraints on the

selectionoftopicsandmethodologiesandthedisseminationofresearchresults.

Soweestimateamodelwhereacademicresearchoutputisaquadraticfunctionofthe

degree of collaboration with industry. Additionally, we make use of the variables

described in theprevious section to control for other factors, other than thedegreeof

collaboration,whichmayalsoaffectresearchoutput.Weinclude,forexample,theability

ofhavingraisedEPSRCresearchfunds,whichproxiesforotherresourcestheresearcher

may have available.We also incorporate patent indicator variables to account for the

existence of other constraints on the scope and/or in the dissemination of research

results.Wealso includepastacademicoutput toaccount forother factorsaffecting the

poolofideasoftheresearcherandherabilitytotransformthemintopapers.Toproxyfor

seniorityandtheexistenceofothertimeconstraints,wealsoincludeheracademicrank.

Accordingly,weformulatethefollowingempiricalmodel:

,

where yit stands for academic i's research output at time t either countit, co‐author

weightedcountit,,appliedcountit,orbasiccountit ;hfit‐1istheindicatorvariablehadsome

EPSRC fundingit‐1; finit‐1 is the fraction of EPSRC funding with industryit‐1; pit is the

indicatorvariablepatentitforhavingfiledatleastonepatentattimet;and,xit‐1isavector

15

ofothertime‐variantindividualcharacteristicsincludingAcademicRankit‐1andyear.All

theindependentvariablesexceptthepatentsarelaggedbecauseofthepublicationlead

time.Theerrortermcontainstwosourcesoferror:theacademici’'sfixedeffecttermi,

andadisturbancetermvit.Sincethedistributionsarehighlyskewed,wetakelogarithms

of both the research output and degree of industry collaboration variables. As these

figurescontainzerovalues,weaddtheunitbeforewetakelogarithms.

4.2. Empiricalmethods

The presence of time‐invariant individual factors, i, in the error term produces

correlation among the individual errors across different periods of time, making

Ordinary Least Squares OLS inefficient and yielding incorrect standard errors. In

addition,asexplainedintheconceptualframework,thereareinherentcharacteristicsof

theresearchers e.g.,ability)thatmakethembothmorelikelytopublishandmorelikely

tofindgoodpartnersand,therefore,tohavehigherdegreesofcollaboration.Thiscreates

problemsofendogeneity forourmainvariablesof interest.Thus,we firstestimateour

modelusingaGeneralizedLeastSquareswithafixedeffectsestimator GLSFE .

Butstill,correctingforfixedeffectsalonemaynotbesufficienttoaddressthebiases

introducedby theselectionmechanismsdescribed in theconceptual framework.There

are time‐variant unobserved individual characteristics, such as becoming more

industry‐savvy, more output‐driven, and/or more selective. These time‐variant

individualunobservedtraitsmayaffectbothresearchoutputanddegreeofcollaboration

causing the latter independent variable to be again endogenous. Fixed Effects based

methodsarenotsufficientindealingwiththissourceoftime‐variantendogeneity.Thus,

to obtain consistent estimates of the coefficients of interest, we use an estimator that

instrumentsourcollaborationmeasureswithinstrumentalvariablesthataffectindustry

collaboration but not research output directly GLS FE IV . Below, we discuss which

instrumentsweuse.

GLS models, though, cannot correct for the error autocorrelation created by the

inclusion of past research outputs i.e., lagged dependent variables as explanatory

variables.Therefore,wealsoestimateadynamicGeneralizedMethodofMoments GMM

panel data model: the Arellano‐Bond GMM estimator GMM AB Arellano and Bond,

1991; Blundell and Bond, 1998 . This estimator transforms the model into first

differences and eliminates the individual effects – and, thus, the cause of the

autocorrelationacrosstimeperiods.Laggeddependentvariablesareinstrumentedwith

notonlytheexogenousvariablesdescribedinthenextsubsectionbutalsowithdeeper

16

lagsoftheindependentanddependentvariableswhichareremoteenoughinthepastso

that their correlationwith current publications has been dissipated. Indeed, wemake

sure that theGMMmodels satisfy both theAutocorrelation test and the Sargan test of

over‐identifyingrestrictions.13 Inourcase,therequireddepthofthelagsisthreeperiods,

whereasthefundinghistorygoesbackfiveyears.Tomakesurethatthisisnotthecause

offurtherhiddenautocorrelation,andasarobustnesscheck,wealsoestimateourmodel

usingafundingstockvariablebasedononlytwoyearsoffunding,asopposedtofive.

4.3. Instrumentalvariables

We instrument the fraction of EPSRC funding with industry variable using the

economicactivityoftheareaandtheoverallshareofindustryfundingofthedepartment.

Economic activity of the area is approximated by the yearly number ofmanufacturing

firms, as listed in theCOMPUSTATdatabase, in theownandadjacentpostcodesof the

universitywheretheacademicworks.Theshareoffundingfromindustryreceivedbythe

whole department is obtained from Research Assessment Exercise RAE data, which

provideinformationontheamountofresearchfundsreceivedbyeachdepartmentinthe

UK, decomposed by source public, private, and other funding for the years 1993 to

2007. We also instrument the variable had some EPSRC funding, using the aggregate

amountoffundingreceivedbythedepartment,basedonthesameRAEdata.

Our instruments for thedegreeofcollaborationassumethat localeconomicactivity

andtheoverallindustryinvolvementofthedepartmentdonotaffectindividualresearch

outputbutdohaveanimpactonthe individual’sopportunitytocollaboratewithfirms.

Similarly,totalfundingofthedepartment,ourinstrumentforhadsomeEPSRCfunding,

should not affect individual research output but should have an impact on the

individual’s ability and opportunity to obtain funds. Our instruments are jointly

significantinthefirststageregressions.Theresiduals‐basedSmith‐Blundelltestrejects

theexogeneityofourcollaborationvariables.Subsequently,weusetheSargan/Hansen’s

statistictotestthatourinstrumentssatisfytheover‐identifyingrestrictions.Noticethat,

giventhatsomeofourinstrumentswereonlyavailablefrom1993onwards,thenumber

ofobservationsisreducedintheregressionswithinstruments.

13 The autocorrelation test of the Arellano‐Bond estimator rules out that the residuals’ dynamic structure is asourceofautocorrelationandisthusanignoredcauseofbiasoftheestimates.TheSargantestassumesthatthemodelisidentifiedandteststhevalidityoftheover‐identifyingrestrictions;inourcasethatthedepthofthelagsofthedependentandotherregressorsusedasinstrumentsissufficienttoruleouttheirendogeneity.

17

5. EmpiricalResults

In this section we present our estimates of the impact of the degree of industry

collaboration.Wefirstintroduceourmainresultsandperformrobustnesschecks.Finally,

weanalyzewhethertheresultsdifferacrossobservedcategoriesofresearchers.

5.1. MainResults

Table3reportsthebasicestimatesofresearchoutput,measuredasthenormalcount

ofpublications.Column1displaystheestimatesofthenon‐instrumentedGLSwithfixed

effects GLSFE model.Column2showstheestimatesofourbenchmarkmodel,theGLS

with fixed effects and instrumental variables GLS FE IV . Column 3 adds one‐ and

two‐year lagged counts of publications as explanatory variables GLS FE IV lags .14

Column4uses theArellano‐BondGMMmodel,with laggedendogenousandexogenous

variablesandyeardummiesasinstruments GMM‐AB .

At the bottom of the table, we include goodness of fit statistics. For each GLS

specification,wereporttheR2andtheF‐statisticassociatedwiththejointsignificanceof

all regressors and the joint significance of the instruments. The null of joint

non‐significance is rejected in all themodels. For the GMMmodels, we report i the

Wald Chi2 tests, which reject the joint non‐significance of the regressors; ii the

Sargan/Hansen tests,which are insignificant, suggesting that themodels do not suffer

fromover‐identification, and iii theArellano‐Bond tests,whichdonot reject thenull

thatthereisanabsenceofthird orhigher ordercorrelationofthedisturbancetermsof

ourspecifications,whichisrequiredfortheconsistencyoftheseestimates.

We proceed by reporting the effects of funding and degree of collaboration.Notice

thatthehadsomeEPSRCfundingvariableallowsustocomparethepredictednumberof

publicationsforanydegreeofcollaboration includingzero tothepredictednumberof

publicationsforaresearcherwithoutfunding.Intotal,wehavea“baseline”productivity

prediction,i.e.,theexpectednumberofpublicationsforanacademicwhodoesnothave

anyfunding hf 0andfin 0intheequationinsection4.1 ,anadditionaleffectforthose

thathavenon‐collaborativefunding hf 1andfin 0 andanothereffectemanatingfrom

thedegreeofcollaboration hf 1and fin 0 .Asshown inTable3,all the fundingand

degreeofcollaborationcoefficientshavethesamesign andareallsignificant inallthe

specifications.Thus,wefirstexplaintheresultsusingthebenchmarkspecification.Then

14 Although the GLS IV estimator does not correct for the autocorrelation created by the endogeneity of laggedpublications, we include this specification to compare the resulting coefficients with those obtained using theGMM‐ABestimator.

18

we compare the magnitudes of the coefficients of the benchmark to those of the

non‐instrumentedspecificationtoshedsomelightonwhichoftheselectionmechanisms

describedintheconceptualframeworkisconsistentwiththeempiricalfindings.

Effectoffundinganddegreeofcollaborationwithindustry.Theantilogoftheconstant

termminusone,whichinthebenchmarkspecificationincolumntwoisequalto0.60,is

the baseline productivity prediction, i.e., the expected number of publications for an

academicatthelowestrank lecturer whodoesnothaveanyfundingorpatents.

Turning to our measure of funding, we find that, consistent with our conceptual

framework, the positive and significant coefficient for the variable had some EPSRC

fundingt‐1 β1 in the equation in section 4.1 corroborates that non‐collaborative

research funding enhances research output. The coefficient implies that for a lecturer

withnopatentsthemarginaleffectofhavingnon‐collaborativeEPSRCfundingcompared

tonothavinganyEPSRCfundingatallisequalto0.35 additional publications.15

ThelinearcoefficientofthefractionofEPSRCfundingwithindustryvariable β2in

theequationinsection4.1 ispositiveandsignificant 0.925 andthecoefficientofthe

quadratic term β3 in the same equation is negative and significant ‐1.710 . These

results indicatethattheeffectof thedegreeof industrycollaborationonthenumberof

publications has an inverted U‐shape. According to the estimated curve, the fraction

EPSRC of funding with industry that would result in the maximum number of

publications is 0.31.16 Thus, in the range of 0% to 31%, increasing the fraction of

collaborative EPSRC funding results inmore publications, but beyond that threshold a

higherfractionisassociatedwithadecreasingnumberofpublications.Inotherwords,all

else equal, researchers who have approximately one third of their EPSRC funding in

collaborationwiththeindustryachievethehighestlevelofresearchoutput.

Figure 2 shows the impact of funding and the degree of industry collaboration on

publications for the benchmark specification GLS FE IV as well as that of the other

specifications GLSFE,GLSFE IVwith lags, andGMM‐AB .Using theestimatesofeach

model,weplotthepredictednumberofpublicationsagainstthedegreeofcollaborative

EPSRCfundingforalecturerwithnopatents.ThedegreeofcollaborativeEPSRCfunding

ranges from 0% to 100%, i.e., from no funding involving industry partners all

15 ThismarginaleffectisthedifferencebetweenthebaselinenumberofpublicationswhentheacademichadsomeEPSRCfunding‐theantilogof 0.471 0.199 minusone‐andthebaselinepublicationscalculatedasdescribedabove‐theantilogof0.471minusone.

16 Thisistheantilog minusone ofthefractionxsatisfyingthefirstorderconditionofthenumberofpublications’maximizationproblem,i.e.,x* β2/ ‐2*β3 .

19

non‐collaborative to all funding including industry partners. For the benchmark

specification,wealsoplotthepredictednumberofpublicationsforaresearcherthathas

notreceivedanyEPSRCfunding thepredictednumberissimilarfortheothermodels .

Comparing the GLS IV FE benchmark to the non‐instrumented specifications. For all

specifications,theinterceptandthelinearcoefficientarepositiveandsignificantandthe

quadratictermisnegativeandsignificant.Thus,independentlyoftheestimationmethod

chosen,theeffectofthedegreeofindustrycollaborationonthenumberofpublicationsis

curvilinearandthecurvehasaninvertedU‐shape,asforthebenchmark.Noticeably,the

curveestimatedwithoutinstrumenting GLSFE isflatter,peaksathigherdegreesand,

more importantly, liesbelow theGLS specifications that instrument collaboration GLS

FEIV,GLSFEIVwithlags fordegreesofcollaborationbelow80%andaboveforthose

above80%.Thus,notaccountingfortheendogeneityproblemresultsinanegativebias

forlowandmediumdegreesofcollaborationandapositivebiasforhighdegrees.

Relating these empirical findings to our conceptual framework, we conclude that

whenwe instrument thedegreeof industry collaborationusingmeasures of economic

activity around the academics’ universities, and thus control for supply side

opportunities forcollaboration,weremovepartof thepositivebiasonpublications for

high degrees of collaboration introduced by the presence of a large proportion of

industry‐savvy researchers. These academics manage to select and to be selected by

high‐qualitypartners.Contrarily, thenegativebiasof theun‐instrumentedspecification

observedatthelowerendofthedegreeofcollaborationcouldbeduetothepresenceofa

large proportion of non‐industry‐savvy researchers that have not developed the same

screeningcapacity,andarenotasappealingacademicpartnerstocollaboratewith.

The negative bias for low degrees of collaboration is also consistent with the

existence of less‐able academics becoming less output‐driven and choosing or being

forcedtochoose degreesofcollaborationthataretoolow.Incontrast,thenegativebias

obtained for intermediate degrees of collaboration is inconsistentwith the hypothesis

that talented and output‐driven academics would choose intermediate degrees of

collaboration. The positive bias observed for high degrees of collaboration is also

inconsistentwiththehypothesisthatless‐ableandlessoutput‐drivenacademicschoose

degreesofcollaborationthataretoohigh.

Finally,thepotentialpositivebiasofthehighlyselectiveacademicsatthelowerend

ofthedistributionofdegreesofcollaborationdoesnotexistormustbeovercomebythe

negative bias introduced by the non‐industry‐savvy researchers or the less

20

‐output‐driven,i.e.,theproportionofhighlyableselectiveacademicsinthelowerdegrees

of collaboration appears to be low. At high degrees of collaboration, the plausible

negative bias introduced by lower‐able academics undertaking a high number of

collaborative projects with low publication potential, seems to be dominated by the

positivebiasintroducedbytheindustry‐savvyresearchers.

Effects of past research output: Note that the curve obtained using the GMM‐AB

specificationisaboveallothercurvesforalldegreesofcollaboration.Thisisbecause,as

shown in column 4 of Table 3, the GMM‐AB coefficient associated with the previous

year’spublicationsispositive,significant,andlarge.Becausewehavetakenlogarithms,

we can interpret this coefficient as an elasticity. Thus, an increase in the number of

publications in the previous year by 100% i.e., doubling them would result in the

followingyear’sexpectednumberofpublicationsincreasingby82percentagepoints.

As inearlierpapers,ourestimates suggest that there ispersistence inpublications.

Severalexplanationsarepossible.First,therecanbea“Mattheweffect” Merton,1968 ,

which describes the possibility that the work of those with a higher number of

publicationsreceivegreaterrecognitionthanequivalentworkbythosethatpublishless.

Second,ahighernumberofpublicationsreinforcestheattractivenessoftheacademicas

a collaborator, increasing her chances of obtaining more and/or better industrial

partnersand,inturn,higherreturnsonhercollaborationsintermsofresearchoutput.

Still,theGMM‐ABresultsdonotchangequalitativelyourresultsandcorroboratethat

therelationshipbetweenthedegreeofcollaborationwithindustryandresearchoutput

canberepresentedasaninvertedU‐shapeevenwhenwecontrolforthepositiveeffectof

pastresearchperformance.

Effectsoftheotherexplanatoryvariables.Wecontrolforthenumberofpatentsfiled

bytheacademictotakeintoaccountotherconstraintsonthescopeofresearchand/orin

thedisseminationofresearchresultsandtocompareourresultstopreviouspapers e.g.,

Azoulayetal.,2009;Breschietal.,2008 .Inaccordancewiththerecentliterature,having

filedpatentsinthecurrentyear t ispositivelyassociatedwithpublications,bothforthe

GLSIVFEandtheGMM‐ABmodels.Themarginalincreases,however,aresmall.17 Inthe

GLSIVFEmodel,havingfiledpatentsineachoftheprevioustwoyears t‐1andt‐2 is

also positive and significant. The marginal increases associated with patents filed in

17 The associated effect is calculated as the antilog minus one of the coefficient of the indicator variable andrangesfrom0.04to0.06extrapublications.

21

currentand inpreviousyearsareverysimilar, thusnotsuggestingapublicationdelay.

ThesecoefficientsarepositivebutnotsignificantintheGMM‐ABmodel.

Wefurthercontrol foracademicrankbutwefindasignificanteffectofseniorityon

publications only for the non‐instrumented model. The effect of seniority in the

instrumental variables’ regressions is absorbed by the instrumented funding variable.

Senioritymaythusbebetteratexplainingaccesstofundingthanpublicationcounts.

5.2. Robustnesschecks

In Table 4, we reproduce the results of our benchmark model using different

measuresofresearchoutputandcollaborationandabalancedsampleofacademics.

Co‐author‐weighted countof publications.The first column shows the impact of the

fractionofcollaborativeEPSRCfundingwhenpublicationsareweightedbythenumber

ofco‐authors.Thebaselinenumberoftheco‐authorweightedpublicationcountis0.33.

ThecoefficientsofthelinearandquadratictermsofthefractionofEPSRCfundingwith

industry variable are significant. The effect of thehad someEPSRC funding variable is

insignificant,suggestingthatfundingmayincreasethenumberofpublicationssimplyby

increasing the sizeof the teams.This is further suggestedby thenegative effect of the

professordummy,asthesemayprimarilybenefitfromlargerlabsthroughco‐authorship.

TheeffectofpatentsispositiveandsignificantasinthebenchmarkregressioninTable3.

Count of basic and appliedpublications. Columns2 and3 decompose the effects by

researchorientation.Wereporttheestimatesoftheimpactofthedegreeofcollaboration

on the countof applied “technology” andbasic “science” publications.Thebaseline

numberof appliedandbasicarticles is, respectively,0.33and0.11.Thus, theexpected

numberofbasicpublicationsislowerthanthatforappliedpublications.Theexistenceof

EPSRC fundingpositively impacts thenumber of basicpublications, but the fractionof

collaborative EPSRC funding does not. For applied research, the linear and quadratic

coefficients associated with industry collaboration variables are instead significant,

similartothebenchmarkregressioninTable3.Thisisduetothefactthatweconsider

thefieldofengineering,wherethemajorityofpublicationsareclassifiedas“applied.”

The effectof patents alsodiffersby typeof research.Our results in thebenchmark

specificationinTable3showthathavingfiledapatentinthecurrentandineachofthe

twopreviousyearssignificantlyincreasestheoverallnumberofpublications.Columns2

and3inTable4showthatwhenseparatingtheeffectforappliedandbasicpublications,

all thecoefficientsofhaving filedpatentsretainthepositivesign.Thereare interesting

22

differences, however, in terms of magnitude and significance. We find a significant

positive contemporaneous effect of patenting on basic publications and a delayed

significantpositive effect onappliedones. Indeed,while basic researchmayproduce a

varietyofcomplementaryresearchoutputsthatcanresultinpublicationsandpatents,it

maybedelayingthepublicationofmoreapplied,technically‐oriented,researchpapers.

Two‐yearbasedfundingstock.Theestimatesincolumn4useavariationofthemain

explanatoryvariables.Here,thevariableshadsomeEPSRCfundingandfractionofEPSRC

fundingwithindustryincludethestreamoffundsreceivedinthelasttwoyearsonly as

opposedtothelastfive .Althoughthischoiceisalessaccuratereflectionofthefunding

profileoftheacademic,itdealsbetterwithpotentialautocorrelationissues.Additionally,

basingourmeasureonashorterwindowallowsyoungerandmoremobileresearchersto

enterthesample.Allcoefficientsofinteresthavethesamesignandsimilarmagnitudeas

inthemainfive‐yearstockregressions.Thedegreeofcollaborativefundingresultingin

themaximum number of publications is lower than that of the benchmark regression

0.17asopposedto0.31 .Thismaybeduetothefactthat fundedprojectsusually last

longer than twoyears.Thepositive, long‐termeffectsof collaborationmaynotbewell

captured and the degree‐maximizing output is smaller. Instead, themodel attributes a

verylargepartofthevariationinpublicationstothevariablehadsomeEPSRCfunding.

Balanced sample. The specification in column 5 is estimated using only those

researchersthatcanbeobservedforthefulllast15yearsofoursample,sothatweare

able tobuild the five‐year funding and industry collaboration variables and estimate a

balanced10‐yearpanel.Thisspecificationenablesustoexplorewhetherthefull‐sample

estimateshavebeen significantlyaffectedbyattrition.Again, all coefficientsof interest

have the same sign and similarmagnitude as in the full‐sample regressions,whichwe

interpretasbeinganindicationthatattritionhasnotcausedimportantbiasesinthemain

estimates.Thedegreeresultinginmaximumnumberofpublicationsisalsoverysimilar.

5.3. Resultsbycategoriesofacademics

In Table 5, we test how our benchmark model results differ across categories of

academicsbyinteractingourmainexplanatoryvariables hadsomeEPSRCfundingand

fraction of EPSRC fundingwith industry with different group‐indicator variables. For

each categorization, the group indicator variable takes the value 0 if the academic

belongstotheso‐calledreferencegroupandthevalue1 if theacademicbelongstothe

non‐reference group. Thus, the main coefficient of a given regressor for instance,

23

fraction ofEPSRC fundingwith industryt‐1 reflects its effect for those in the reference

group.Theeffectforthoseinthenon‐referencegroupisobtainedbyaddingtothemain

coefficientthecoefficientassociatedtotheinteractedterm groupindicator*fractionof

EPSRCfundingwithindustryt‐1 , ifsignificantlydifferentfromzero.Asexplainedbelow,

thetrade‐offbetweenindustrycollaborationandpublicationsexistsforallthecategories

of researchersanalyzed.That is, theeffectof the linear termof thevariable fractionof

EPSRC with industry is positive and that of the quadratic term is negative for all

specifications,both for the referenceand thenon‐referencegroups.Themagnitudesof

theeffects,however,differ.

The first column distinguishes between academics who are high collaborators in

terms of having an above themedian share of publications co‐authoredwith industry

referencegroup fromthosethathaveanaveragebelowthemedian.Themaineffects

for the linear and quadratic terms of the variable fraction of EPSRCwith industry are

larger than those in thebenchmarkcase inTable3.Thus the relationshipbetween the

degree of collaboration and publications for high‐collaborators is characterized by a

moreconcavecurve thantherelationshipcorresponding to thebenchmark’sestimates.

The interaction terms’ estimates are significant and of the opposite sign, albeit of a

smaller magnitude, than the main effects. This indicates that the relationship for the

group of low‐collaborators is still an inverted U‐shape, albeit less concave. Therefore,

occasional collaboratorsalsoexperiencegainsand losseswhenvarying theirdegreeof

collaboration, but the effects are weaker. Indeed, the breadth of ideas, the funding

obtained,aswellas theconstraints imposed,seemrelativelymore important forheavy

collaborators.

Thesecondcolumndistinguishesacademicswithanabove‐the‐medianpercentageof

averagelifetimeofEPSRCfundingwithindustry referencegroup fromthosebelow.The

magnitude of the main effects on publications is larger than those in the benchmark

model and is also larger than those in column1. This suggests that high collaborators

basedonlifetimefundingexhibitarelationshipbetweenthedegreeofcollaborationand

publications that is evenmore sensitive to changes in thedegreeof collaboration than

the relationship of high‐collaborators based on joint publications. The correction

estimatesforlow‐collaboratorsasperthismeasurearenotsignificant.

Thethirdcolumnseparatesacademicswithhighlevelsoffunding,i.e.,thoseabovethe

median in the amount of EPSRC funding received in the last 5 years, from thosewith

lower levels of funding. Again themain coefficients of the funding variables are larger

24

thanthoseofthebaselinemodelinTable3.Inaddition,thecorrectioncoefficientsforthe

non‐referencegrouparenotsignificantlydifferentfromzero.

Incolumn4,wedistinguishbetweenacademicsworkinginaRussellgroupuniversity

fromthosethatdonot.Wefindsimilarcoefficientstothoseinthebenchmarkmodeland

no significant difference between researchers in the group of well‐known research

intensive institutions from those in lesser‐known and less‐funded institutions. Lower

levelsofcorefundingmayforcesmaller institutionstorelyrelativelymoreonexternal

grants Perkmannetal.,2013 andresultinhigherdegreesofengagementwithindustry

D’Este and Patel, 2005 . But, the trade‐offs of industry collaboration in terms of

academicoutputappeartobesimilarforacademicsatverydifferentkindsofinstitutions.

Lastly,column5distinguishesbetweenacademicsthatareofaloweracademicrank

fromthoseholdingtherankoffullprofessor.Thetrade‐offassociatedwiththedegreeof

collaborationmayhavebeenlesspronouncedforsenioracademics.Extensiveexperience

and consolidated networks could make the new insights and the additional funds

acquiredthroughcollaborationrelativelylessrelevant,butatthesametimeconstraints

may also be less important. Young researchers, instead, are at a crucial point of their

careers, and their research output is expected to be relatively more sensitive to

collaboration Dasgupta and David, 1994 . According to our estimates, though, a

professordoesnotexperienceasignificantlydifferent impactof thedegreeof industry

collaborationcomparedtosomeoneataloweracademicrank.

Overall,we conclude that although there are significant differences, the curvilinear

effectofthedegreeofindustrycollaborationholdsfordifferentcategoriesofacademics.

Weinterpretthisasevidenceoftherobustnessofourresults.

6. DiscussionandConclusion

The effect of research collaboration on publication outcomes has received little

attention in the academic literature to date,which has primarily focused on academic

patenting and spin‐off formations as channels of interaction between science and

industry. Many authors have argued, though, that research collaborations, contract

research,andconsultancyarefarmoreimportantchannelsofknowledgetransfer.They

are,however,moredifficult tomeasureempiricallyandevenmoredifficult tocompare

across institutions and time, which may explain why the literature has paid scant

attentiontothesemorecollaborativeformsofuniversity‐industryinteractions.

25

This paper uses homogeneous information on the collaborative grants awarded by

theEPSRC,byfarthemostimportantfundingsourceofresearchinengineeringsciences

intheUK,tomeasureuniversity‐industrycollaborationovera20‐yearperiod.Weshow

that the effect of collaboration depends on the share of projects undertaken in

collaborationwithindustry,i.e.,onthedegreeofcollaboration.Ourresultsindicatethat

thenumberofpublicationsincreasesbothwiththepresenceofEPSRCfundingandwith

the fraction of EPSRC funding in collaborationwith industry, but only up to a certain

point.Fordegreesofcollaborationabove30%–40%,researchoutputdeclines.

These results confirm the expectations of our conceptual frameworkwhich argues

that the degree of collaboration‐publication relationship could be described by an

invertedU‐shapedcurve.Indeed,theformationoflinkswiththeprivatesectormayboost

researchoutputbecausecollaborationcanprovidenewideasandadditionalfunding.But,

high degrees of collaboration can also damage research output, as research ideasmay

thenbeoflowervalue,industrymayimposenon‐disclosureclausesorbecauseextensive

collaborationcouldreducethetimetodoresearchandcauseattentionproblems.

Ourresultsmayprovideanexplanationforsomeofthe apparentlymixed resultsin

theliterature,whichhaseffectivelyfocusedonlinearrelationships.Ontheonehand,they

might explain the positive effects documented in studies that investigate forms of

collaborationthatrequirelittletonodirectinteraction e.g.,contactthroughtechnology

transferofficesandtradefairs,asinHottenrottandLawson,2014 .Academicpatenting

can also be viewed as collaboration requiring low levels of interaction Agrawal and

Henderson, 2002 . The evidence of a positive relationship between patenting and

publications Azoulay et al., 2009; Breschi et al., 2008; Fabrizio and DiMinin, 2008;

Stephanetal.,2007;vanLooyetal.,2005 isthenconsistentwiththeideaofpatenting

beingontheincreasingpartoftheinvertedU‐shapedcurve.

But,ontheotherhand,ourresultscanalsoexplainthenegativeeffectsdocumentedin

other studies. Previous research that investigates the effects of forms of collaboration

thatrequiresubstantialinteraction,asforexampleinacademicstart‐ups,findanegative

effect.TooleandCzarnitzki 2010 showthatUSacademicsthatreceivefundingtostart

orjoinfor‐profitfirmsaremoreproductivethantheirpeers,butthattheyproducefewer

publicationsafterreceivingthegrant.Goldfarb 2008 tracksasampleof221university

researchers funded by theNASA and concludes that researchers repeatedly funded by

theNASAexperiencedareductioninacademicoutput.Inthesecases,theeffectislikely

tohavebeenthatassociatedtothedecreasingpartoftheinvertedU‐shapedcurve.

26

Our results are robust to various measures of academic research output and to

various subsamples of academics. Nevertheless, some remarks are in order. Basic

research output is positively affected by the presence of EPSRC funding but not

significantly affected by the degree of industry collaboration. Conversely, for applied

publications we find no significant effect of the presence of EPSRC funding but a

significant invertedU‐shapedeffectof thedegreeof industrycollaboration. Indeed, the

effectsofcollaborationidentified intheconceptual frameworkareespecially important

for applied research. The availability of financial resources, one of the benefits of

collaboration, is key for applied research programs.18 Research ideas arising from

collaborationarealsomorelikelytobeturnedintoappliedresearchpapers.Atthesame

time, publication constraints should especially affect applied publications as applied

research canbepublishedaspatentsorpublications, resulting inpotentialpublication

delays PerkmannandWalsh,2009 .

Akeychallengeinouranalysisisthattheobserveddegreesofcollaborationarenot

exogenous,butaretheresultofindividualandbilateralchoicesinatwo‐sided“market”

ofacademicsandfirms.Ourconceptual frameworkshowsthatthedirectionofthebias

onemightincur,ifselectionissuesareignored,dependsonwhichmechanismisatwork.

Our empirical analysis compares the results of our instrumental‐variable benchmark

specification to those of a non‐instrumented regression to shed some light on which

mechanism is the most important in our data. We document a negative bias for low

degreesofcollaborationandapositivebiasforhighdegrees.Thisisconsistentwiththe

presence of a large proportion of “industry‐savvy” researchers at the high‐end of the

range of the degrees of industry collaboration. Indeed, this presence shall introduce a

positivebias inpublicationsbecause industry‐savvyresearchersendupbeingmatched

withmoreproductivefirmpartners,resultinginmoreproductiveresearchprojects.

Ourresultsalsobolsterempiricalevidencefromprevioussurveysandcross‐sectional

studiesontheeffectsofcollaborativeresearchfundingonacademicoutputbyshowing

that these results hold for a large longitudinal sample. Even after controlling for

endogeneity, we find supportive evidence of the positive impact of the existence of

collaboration,asinGulbrandsenandSmeby 2005 .Thenegativeeffectofhighdegrees

ofcollaborationisalsoconsistentwithothersurveyresults Blumenthaletal.,1996 and

cross‐sectionempiricalevidence Manjarres‐Henriquezetal.,2009 .

18 Financialrewards,however,mightalsohaveapositiveimpactontheproductionofbasicresearchbecausebasicand applied research efforts are complementary Thursby et al., 2007 or because they induce a selection ofriskierandmorebasicresearchprogramms Banal‐EstañolandMacho‐Stadler,2010 .

27

Wealsofindadirect,positiveeffectofpatentingonpublications,justasintheearlier

literature.Thecontemporaneouseffectofpatentdisclosureissimilartotheoneofpast

patents,givingnoevidenceofa“secrecy”effect.Yet,whenwedistinguishbetweentypes

of research, we find a stronger contemporaneous effect of patents on publications in

basic science journals. This may explain the positive correlation found in papers that

analyze publication and patenting activity of researchers in basic sciences, e.g.,

life‐science Azoulayetal.,2009;Breschietal.,2008 andthelackofcontemporaneous

correlationfoundinpapersthatanalyzeappliedsciences,e.g.,engineering Agrawaland

Henderson, 2002 . We do, however, find a delayed positive effect of patenting for

publications in applied journals, suggesting that applied research may suffer from

secrecy.PerkmannandWalsh 2009 indeedarguethatmoreappliedprojectsaremore

likelytobeaffectedbysecrecybecauseoftheirimmediatecommercialviability.

In terms of policy or managerial implications, our findings suggest that program

interventions encouraging academic researchers to collaborate with industry may be

beneficial.Amoderatedegreeofindustrycollaborationnotonlyfacilitatesthetransferof

knowledge and accelerates the exploitation of new inventions, but it also increases

academic research output. Our results point at the disadvantages of academic policies

that favor the separation of tasks at the university. If some facultymembers focus on

researchwhile others performother activities such as collaboration and development,

scientific output might suffer. Indeed, according to our estimates, two academics

collaborating moderately degree of 30–50% would publish more than a

non‐collaboratingandafullycollaboratingonecombined degreesof0%and60–100% .

On theotherhand, our results also indicate that there aredegreesof collaboration

thatmaybeexcessiveinthesenseofbeingdetrimentalintermsofresearchproductivity.

Several factors point to an increase in the degrees of industry collaboration in recent

times Stephan, 2012 . As the degrees of involvement with industry increase, more

academicsmay find themselveson thedecreasingpartof the curve. Indeed, academics

might start pursuing research lines that no longer result in breakthrough discoveries

resultinginfewerpublications.Inaddition,highdegreesofcollaborationcannegatively

affect material and data exchange between academics and thus be damaging to the

academiccommunityasawhole Stephan,2012 .But,highdegreesofcollaborationcan

alsobringgainsintermsofpatentingorbetteremploymentprospectsforgraduates.

Ourscanonlybea firststep in theanalysisof theeffectsof thevariouschannelsof

knowledge transfer.Wehad to limit our analysis to research collaborations sponsored

28

throughpublicfunding,whichcanonlyproxyfortheextentofthecollaborationactivities.

Including private partners in these grants, though, is highly correlated with obtaining

direct funding from the industry (Meissner, 2011). We therefore expect that the

curvilineareffectwouldremainunhinderedif,insteadofusingourmeasureofdegreeof

industry collaboration, we were to use the overall degree of engagement. In terms of

magnitudes, though, the curve may peak at higher degrees of collaboration. First,

mechanically, ifdirectfundingisaddedinthenumeratoranddenominator,thefraction

increases.Second,becauseofthepositivecorrelation,theresearcherswithonethirdof

theirEPSRCfundswithindustryshallhaveasubstantialamountofdirectfunding.

With more comprehensive and homogeneous information, we could also make

comparisons between the effects of research collaboration and other channels of

knowledge transfer such as consultancy and patents. In our sample, research

collaborationshaveastrongerimpactthanpatents.Itmightalsobeofinteresttotackle

interactionsbetweenthedifferentchannels.Weknowlittleaboutwhethercollaboration

channels complementor substitute eachother.Consultancy, for example,mighthavea

positive effect on researchoutput if andonly if it is complementedby collaboration in

research.Ofcourse,thisisonlyaconjectureandachallengingtaskforfutureresearch.

Acknowledgements

We thank the editor,AshishArora, and twoanonymous referees for their constructive

feedback.WealsothankMichaelBen‐Gad,InesMacho‐Stadler,AlistairMcGuire,Gabriel

Montes‐Rojas, Markus Perkmann, Dan Peled, Jo Seldeslachts, Anna Torres, seminar

participantsatCityUniversityLondon,Sheffield,Munich,Ingenio‐CSIC Valencia ,DRUID

conference Copenhagen ,JEI Reus ,ASSETconference Istanbul andMOVEworkshop

onR&DandTechnologyTransfer Barcelona forhelpfulcomments,andtheEPSRC,the

NSFandtheEPOforprovidingthedata.WethankKamilDabrowskiandXeniaRadufor

theirexcellentresearchassistance.TheresearchbenefittedfromaPumpprimingfundat

City University and the ESRC research grant RES‐000‐22‐2806. Albert Banal‐Estañol

acknowledges financial support from the Ministerio de Ciencia y Tecnologia

ECO2010‐15052,ECO2013‐43011‐PandRamonyCajal ,andCorneliaLawsonsupport

fromtheEuropeanCommission FP7#266588andDIMEDoctoralMobilityFellowship

andaCollegioCarloAlbertogrant.Anearlierversionofthispaperwascirculatedunder

thetitle“TheImpactofIndustryCollaborationonResearch:EvidencefromEngineering

AcademicsintheUK“.

29

References

Agrawal,A.andR.Henderson,2002,“Puttingpatentsincontext:ExploringknowledgetransferfromMIT”,ManagementScience48,44–60.

Arellano,M.andS.Bond,1991,“Sometestsofspecificationforpaneldata:MonteCarloevidenceandanapplicationtoemploymentequations”,ReviewofEconomicStudies58,277–297.

Arora,A.,P.DavidandA.Gambardella,1998,“ReputationandCompetenceinPubliclyFundedScience:EstimatingtheEffectsonResearchGroupProductivity”,Annalesd’

EconomieetdeStatistique49–50,163–198.

Audretsch,D.andP.Stephan,1996,“Company‐ScientistLocationalLinks:TheCaseofBiotechnology”,TheAmericanEconomicReview86,641–652.

Azoulay,P.,W.DingandT.Stuart,2009,“TheImpactofAcademicPatentingontheRate,Quality,andDirectionof Public Research”,TheJournalofIndustrialEconomics,57 4 ,637–676.

Baldini,N.,2008,“Negativeeffectsofuniversitypatenting:Mythsandgroundedevidence”,Scientometrics,75 2 ,289–311.

Banal‐Estañol,A.,andI.Macho‐Stadler,2010,“ScientificandCommercialIncentivesinR&D:Researchvs.Development?”JournalofEconomicsandManagementStrategy,19 1 ,pp.185–221.

Banal‐Estañol,A.,Macho‐Stadler,I.andD.Pérez‐Castrillo,2013,“ResearchOutputfromUniversity‐IndustryCollaborativeProjects”,EconomicDevelopmentQuarterly271 ,pp.71‐81.

Banal‐Estañol,A.,I.Macho‐StadlerandD.Pérez‐Castrillo,2014,““EndogeneousMatchinginUniversity‐IndustryCollaboration:TheoryandEmpiricalEvidencefromtheUK”,workingpaperUniversitatPompeuFabra.

Benavente,J.M.,G.Crespi,L.FigalGaroneandA.Maffoli,2012,"Theimpactofnationalresearchfunds:AregressiondiscontinuityapproachtotheChileanFONDECYT",ResearchPolicy41,1461–1475.

Blumenthal,D.,N.Causino,E.G.CampbellandK.S.Louis,1996,“RelationshipsbetweenAcademicInstitutionsandIndustryintheLifeSciences — AnIndustrySurvey.”,

NewEnglandJournalofMedicine,334 6 ,pp.368–73.

Blumenthal,D.,M.Gluck,K.S.Louis,M.A.StotoandD.Wise,1986,“University‐industryResearchRelationshipsinBiotechnology:ImplicationsfortheUniversity”,Science13June,232,4756,1361–1366.

Blundell,R.andS.Bond.1998,“Initialconditionsandmomentrestrictionsindynamicpaneldatamodels”,JournalofEconometrics87,11–143.

30

Breschi,S.,F.LissoniandF.Montobbio,2008,““Universitypatentingandscientificproductivity.AquantitativestudyofItalianacademicinventors”,EuropeanManagementReview5,91–110.

Carayol,N.,2003,“Objectives,agreementsandmatchinginscience–industrycollaborations:reassemblingthepiecesofthepuzzle”,ResearchPolicy32,887‐908.

Cohen,W.M.,R.R.NelsonandJ.P.Walsh,2002,““LinksandImpacts:TheInfluenceofPublicResearchonIndustrialR&D”,ManagementScience48,1–23.

Czarnitzki,D.,C.GrimpeandA.Toole,2014,““DelayandSecrecy:DoesIndustrySponsorshipJeopardizeDisclosureofAcademicResearch?”IndustrialandCorporateChange,forthcoming.

D’EsteP.andP.Patel,2005,“University–industrylinkagesintheUK:Whatarethefactors

underlyingthevarietyofinteractionswithindustry?” ResearchPolicy36,1295–1313.

Dasgupta,P.andDavid,P.A.,1994,“Towardaneweconomicsofscience”ResearchPolicy23 5 ,487–521.

EuropeanCommission,1995,“GreenPaperonInnovation”,Luxemburg.

Fabrizio,K.andA.DiMinin,2008,“CommercializingtheLaboratory:FacultyPatentingandtheOpenScienceEnvironment”,ResearchPolicy37,914–931.

Florida,R.andW.M.Cohen,1999,“EngineorInfrastructure? TheUniversityRoleinEconomicDevelopment”In:Branscomb,L.M.,Kodama,F.,Florida,R. Eds. ,IndustrializingKnowledge:University—IndustryLinkagesinJapanandtheUnited

States.MITPress,London,589–610.

Geuna,A.andL.Nesta,2006,“Universitypatentinganditseffectsonacademicresearch:TheemergingEuropeanevidence”,ResearchPolicy,35 6 ,790–807.

Godin,B.,1996,“Researchandthepracticeofpublicationinindustries”,ResearchPolicy25,587–606.

Goldfarb,B.,2008,“TheEffectofGovernmentContractingonAcademicResearch:DoestheSourceofFundingAffectScientificOutput?,ResearchPolicy37,41‐58.

Gulbrandsen,M.andJ.C.Smeby,2005,“IndustryFundingandUniversityProfessors’ResearchPerformance”,ResearchPolicy34,932–950.

Hall,B.,2004,“University‐IndustryResearchPartnershipsintheUnitedStates”,inContzen,Jean‐Pierre,DavidGibson,andManuelV.Heitor eds. ,RethinkingScienceSystemsandInnovationPolicies,PurdueUniversityPress.

Henderson,R.,A.JaffeandM.Trajtenberg,1998,“UniversitiesasaSourceofCommercialTechnology:aDetailedAnalysisofUniversityPatenting,1965–1988”,Reviewof

EconomicsandStatistics80,119–127.

31

Hottenrott,H.andC.Lawson,2014,"ResearchGrants,SourcesofIdeasandtheEffectsonAcademicResearch",EconomicsofInnovationandNewTechnology,23 2 ,109‐133.

Jacob,B.A.andL.Lefgren,2011,“Theimpactofresearchgrantfundingonscientificproductivity”,JournalofPublicEconomics95:1168–1177.

Laursen,K.,T.ReichsteinandA.Salter,2011,“ExploringtheEffectofGeographicalDistanceandUniversityQualityonIndustry‐UniversityCollaborationintheUK”,RegionalStudies45,507–523.

Laursen,K.,andA.Salter,2004,“Searchinglowandhigh:whattypesoffirmsuseuniversitiesasasourceofinnovation”,ResearchPolicy33,1201–1215

Laursen,K.,andA.Salter,2006,“Openforinnovation:TheroleofopennessinexplaininginnovationperformanceamongUKmanufacturingfirms”,StrategicManagementJournal27,131‐150.

Lawson,C.,2013a,“AcademicPatenting:TheImportanceofIndustrySupport”,JournalofTechnologyTransfer38 4 ,509–535.

Lawson,C.,2013b,“Academicinventionsoutsidetheuniversity:InvestigatingpatentownershipintheUK”,IndustryandInnovation,20 5 ,385–398.

Lee,Y.S.,2000,“Thesustainabilityofuniversity‐industryresearchcollaboration:Anempiricalassessment”,TheJournalofTechnologyTransfer,25 2 ,111–133.

LinkA.N.andJ.T.Scott,2005,“UniversitiesaspartnersinUSresearchjointventures”,ResearchPolicy34,385–393.

Lissoni,F.,P.Llerena,M.McKelveyandB.Sanditov,2008,“AcademicpatentinginEurope:newevidencefromtheKEINSdatabase”,ResearchEvaluation17,87–102.

Manjarres‐Henriquez,L.,A.Gutierrez‐Gracia,A.Carrion‐Garcia,andJ.Vega‐Jurado,2009,“Theeffectsofuniversity–industryrelationshipsandacademicresearchonscientificperformance:Synergyorsubstitution?”ResearchinHigherEducation,508 ,795–811.

Mansfield,E.,1995,“AcademicResearchUnderlyingIndustrialInnovations:Sources,Characteristics,andFinancing”,ReviewofEconomicsandStatistics77,55–65.

Maurseth,P.,andB.Verspagen,2002,“KnowledgeSpilloversinEurope:APatentCitationsAnalysis”,ScandinavianJournalofEconomics,104 4 531–45.

Meissner,C.,2011,“UniversityResearchandIndustryInvolvement.ThreeEssaysontheEffectsandDeterminantsofIndustryCollaborationandCommercialisationinAcademia”,PhDThesis,CityUniversityLondon.

Merton,R.K.,1968,“TheMatthewEffectinScience:theRewardandCommunicationSystemsofScienceConsidered”,Science159,pp.56‐63.

Meyer‐Krahmer,F.,Schmoch,U.,1998,“Science‐basedtechnologiesuniversity–industryinteractionsinfourfields”,ResearchPolicy27,835–852.

32

Mindruta,D.,2013,“ValueCreationinUniversity‐FirmResearchCollaborations:AMatchingApproach”,StrategicManagementJournal34,644–665.

Mohnen,P.andC.Hoareau,2003,“WhatTypeofEnterpriseForgesCloseLinkswithUniversitiesandGovernmentLabs?EvidencefromCIS2”,ManagerialandDecisionEconomics24,133–146.

Mowery,D.,R.Nelson,B.Sampat,andA.Ziedonis,2001,“ThegrowthofpatentingandlicensingbyU.S.universities:anassessmentoftheeffectsoftheBayh–Doleactof

1980”.ResearchPolicy30,99–119.

Narin,F.,G.Pinski,andH.Gee,1976,“StructureoftheBiomedicalLiterature,”JournaloftheAmericanSocietyforInformationScience,25–45.

Nelson,R.,2004,“Themarketeconomy,andthescientificcommons”,ResearchPolicy,33 3 ,455–471.

Ocasio,W.,1997,“Towardsanattention‐basedviewofthefirm,”StrategicManagementJournal18,187–206.

OECD OrganisationforEconomicCo‐operationandDevelopment 2009.Business‐FundedR&DintheHigherEducationandGovernmentSectors,in:OECDScience,TechnologyandIndustryScoreboard2009.Paris:OECDPublishing.

Perkmann,M.,V.Tartari,M.McKelveyetal.,2013,“Academicengagementandcommercialisation:Areviewoftheliteratureonuniversity–industryrelations”,

ResearchPolicy42,423–442.

Perkmann,M.andK.Walsh,2009,“ThetwoFacesofCollaboration:ImpactsofUniversity‐IndustryRelationsonPublicResearch”,IndustrialandCorporateChange18,1033–1065.

Rosenberg,N.,1998,“ChemicalEngineeringasaGeneralPurposeTechnology”,in:Helpman,E. Ed. ,GeneralPurposeTechnologiesandEconomicGrowth.Cambridge:MITPress,167–192.

Rosenberg,N.andR.Nelson,1994,“AmericanUniversitiesandTechnicalAdvanceinIndustry”,ResearchPolicy23,323–348.

Salter,A.andB.R.Martin,2001,“TheEconomicBenefitsofPubliclyFundedBasicResearch:aCriticalReview”,ResearchPolicy30,509–532.

Siegel,D.S.,D.WaldmanandA.Link,2003.“AssessingtheImpactofOrganizationalPracticesontheRelativeProductivityofUniversityTechnologyTransferOffices:AnExploratoryStudy”,ResearchPolicy32,27–48.

Stephan,P.,1996,“TheEconomicsofScience”,JournalofEconomicLiterature34,1199–1235.

Stephan,P.,2012,“HowEconomicsShapesScience”,Cambridge,MA:HarvardUniversityPress.

33

Stephan,P.,S.Gurmu,A.J.SumellandG.Black,2007,“Who’sPatentingintheUniversity?”EconomicsofInnovationandNewTechnology61,71–99.

Stern,S.,2004,“Doscientistspaytobescientists?”ManagementScience50,835–853.

Subramanyam,L,1983,“Bibliometricstudiesofresearchcollaboration:Areview”JournalofInformationScience6,33–38.

Thursby,M.,J.ThursbyandS.Mukherjee,2007,“ArethereRealEffectsofLicensingonAcademicResearch?ALifeCycleView”,JournalofEconomicBehavior&Organization63 4 ,577–598.

Toole,A.,andD.Czarnitzki,2010,“CommercializingScience:IsthereaUniversity‘BrainDrain’fromAcademicEntrepreneurship?”ManagementScience56 9 ,1599–1614

Trajtenberg,M.,R.HendersonandA.B.Jaffe,1997,“UniversityVersusCorporatePatents:aWindowontheBasicnessofInvention”,EconomicsofInnovationandNewTechnology5 19 ,19–50.

vanLooyB.,K.DebackereandJ.Callaert,2006,“PublicationandPatentBehaviourofacademicresearchers:conflicting,reinforcingormerelyco‐existing”,ResearchPolicy35,596–608.

Veugelers,R.andB.Cassiman,2005,“R&Dcooperationbetweenfirmsanduniversities.SomeempiricalevidencefromBelgianmanufacturing”,InternationalJournalofIndustrialOrganization23,355–379.

Russell Group Universities Number of ID*Number of 

ObservationsOther Universities Number of ID

Number of 

Observations

Birmingham, University of 193 1523 Aberdeen, University of 45 358

Bristol University 82 581 Aston University 62 573

Cambridge, University of 191 1541 Bangor University 30 179

Cardiff, University of 99 802 Brunel University 117 810

Durham, University of 43 308 City University, London 62 556

Edinburgh, University of 92 724 Dundee, University of 49 400

Exeter, University of 40 317 Essex, University of 29 295

Glasgow, University of 108 1033 Hull, University of 37 338

Imperial College London 268 2109 Heriot Watt University 141 1119

Kings College London 48 330 Lancaster, University of 26 230

Leeds, University of 165 1212 Leicester, University of 37 245

Liverpool, University of 102 888 Loughborough, University of 236 1847

Manchester, University ofǂ 322 2614 Reading, University of 47 395

Newcastle, University of 143 1188 Salford, University of 99 788

Nottingham, University of 167 1317 Strathclyde, University of 179 1563

Oxford, University of 97 843 Swansea University 94 856

Queen Mary London 82 575

Queens University, Belfast 101 920

Sheffield, University of 176 1293

Southampton, University of 136 1060

University College London 127 1045

Warwick, University of 64 621

York, University of 27 205

* Researchers can belong to more than one university during their career. Therefore the numbers of IDs do not add up to 3,991, the number of unique individuals in our sample.

Table 1: List of Universities

Full Sample (33601 obs) countthad some 

fundingit

fraction of 

funding with 

industryit 

Dependent Variables

countit 1.64 2.68 0 41 1 0.228*** 0.125***

co‐author weighted countit 0.59 0.93 0 12.26 0.916*** 0.208***

0.106***

average impact factorit 0.44 0.72 0 27.07 0.436*** 0.167***

0.100***

applied countit 0.9 1.62 0 24 0.778*** 0.206*** 0.122***

basic countit 0.46 1.52 0 26 0.712*** 0.125*** 0.0216***

Explanatory Variables

had some EPSRC fundingit 0.67 0.47 0 1 0.228*** 1 0.405***

fraction of EPSRC funding with industry it 0.2 0.35 0 1 0.125*** 0.405*** 1

patentit 0.03 0.18 0 1 0.124*** 0.0794*** 0.0577***

lecturerit 0.34 0.48 0 1 ‐0.183*** ‐0.194***

‐0.0998***

senior lecturerit 0.28 0.45 0 1 ‐0.104*** ‐0.0511***

‐0.0204***

readerit 0.1 0.3 0 1 0.0853*** 0.0453*** 0.0276***

professorit 0.27 0.45 0 1 0.242*** 0.228*** 0.109***

Low vs high collaborators in publicationsƚ (27667 obs)

Dependent Variables mean sd min max mean sd Min max

countit 1.43 2.41 0 37 2.22 3.13 0 41 ***

Explanatory Variables

had some EPSRC fundingit 0.59 0.49 0 1 0.73 0.44 0 1 ***

fraction of EPSRC funding with industry it 0.19 0.34 0 1 0.3 0.38 0 1 ***

Low vs high collaborators in fundingƚƚ (23645 obs)

Dependent Variables mean sd min max mean sd min max

countit 1.71 2.62 0 37 2.26 3.2 0 41 ***

Explanatory Variables

had some EPSRC fundingit 0.78 0.42 0 1 0.82 0.39 0 1 ***

fraction of EPSRC funding with industry it 0.07 0.19 0 1 0.5 0.41 0 1 ***

Low vs high recipients of fundingƚƚƚ (28508 obs)

Dependent Variables mean sd min max mean sd Min max

countit 0.97 1.73 0 23 2.45 3.3 0 41 ***

Explanatory Variables

had some EPSRC fundingit 0.38 0.49 0 1 0.89 0.31 0 1 ***

fraction of EPSRC funding with industry it 0.11 0.31 0 0 0.36 0.38 0 1 ***

Type of institution (28508 obs)

Dependent Variables mean sd min max mean sd min max

countit 2.02 3.02 0 41 1.25 2.16 0 33 ***

Explanatory Variables

had some EPSRC fundingit 0.68 0.47 0 1 0.62 0.48 0 1 ***

fraction of EPSRC funding with industry it 0.24 0.37 0 1 0.24 0.38 0 1

Academic rank (28508 obs)

Dependent Variables mean sd min max mean sd Min max

countit 2.86 3.66 0 41 1.36 2.25 0 32 ***

Explanatory Variables

had some EPSRC fundingit 0.84 0.37 0 1 0.59 0.49 0 1 ***

fraction of EPSRC funding with industry it 0.3 0.36 0 1 0.22 0.37 0 1 ***

* p < 0.10, ** p < 0.05, *** p < 0.01.

ƚ High (low) collaborators are those who have an average lifetime collaborative publications with the industry  above (below) the median. Academics with zero publications excluded.

ƚƚ High (low) collaborators are those who have an average lifetime collaborative EPSRC funding with the industry  above (below) the median. Academics with zero funding excluded.

ƚƚƚ High (low) funding receivers are those who have received EPSRC funding in the previous 5 years  above (below) the median.

Table 2: Descriptive Statistics

Correlation with

Low Funding  (12853 obs) High Funding (15655 obs)

Low Collaborators  (13484 obs) High Collaborators (14183 obs)

mean sd min max

Low Collaborators  (11188 obs) High Collaborators (12457 obs)

Russell Group  (18374 obs) Non‐Russell Group (8408 obs)

Professor  (7955 obs) Not Professor (20553 obs)

t‐test 

difference

t‐test 

difference

t‐test 

difference

t‐test 

difference

t‐test 

difference

1 2 3 4

Model GLS FE GLS FE IV GLS FE IV GMM ABƚ

Dependent variable countit countit countit countit

Funding and Collaboration:

had some EPSRC fundingit‐1 0.044*** 0.199** 0.203** 0.066*

(0.013) (0.083) (0.083) (0.035)

fraction of EPSRC funding with industryit‐1 0.373*** 0.925*** 0.897*** 0.286*

(0.099) (0.352) (0.344) (0.167)

fraction of EPSRC funding with industry2it‐1 ‐0.421*** ‐1.710*** ‐1.651*** ‐0.432*

(0.145) (0.655) (0.638) (0.249)

Patent Filed:

patentit 0.026 0.040* 0.040* 0.059**

(0.019) (0.021) (0.021) (0.024)

patentit‐1 0.014 0.037* 0.035* 0.014

(0.019) (0.020) (0.020) (0.025)

patentit‐2 0.037* 0.039* 0.038* 0.017

(0.021) (0.021) (0.021) (0.026)

Academic Rank:

senior lecturerit‐1 0.048*** 0.025 0.022 ‐0.013

(0.014) (0.018) (0.018) (0.009)

readerit‐1 0.093*** ‐0.002 ‐0.007 0.001

(0.022) (0.027) (0.027) (0.020)

professorit‐1 0.109*** ‐0.003 ‐0.009 0.017

(0.025) (0.032) (0.032) (0.019)

Lagged Publications:

countit‐1 0.031*** 0.820***

(0.008) (0.046)

countit‐2 0.003 0.042***

(0.007) (0.015)

Constant 0.479*** 0.471*** 0.453*** 0.087***

(0.014) (0.039) (0.040) (0.022)

Number of observations 33601 28508 28508 26782

Number of ID 3991 3975 3975 3724

Number of instruments 0 3 3 164

F (Joint sig. of instr. in 1st stage) 14.25*** 14.25***

R^2 (overall) 0.247 0.087 0.369

F 19.072*** 8.961*** 8.709***

Wald chi2 12597***

AR(1) test z (p‐value) 0

AR(2) test z (p‐value) 0

AR(3) test z (p‐value) 0.7205

Sargan  test (p‐value) 0.1699

Robust standard errors in parentheses; * p  < 0.10, ** p  < 0.05, *** p  < 0.01.

Table 3: Impact of Industry Collaboration on Research Output ǂ

ƚ Endogenous variables are fraction of EPSRC funding with industry  and lagged count . Lagged endogenous 

and exogenous variables and year dummies are used as instruments.

ǂ The dependent variable, the lagged dependent variables and the fraction of EPSRC funding with industry 

are in logarithms. 

1 2 3 4 5

Model GLS FE IV GLS FE IV GLS FE IV GLS FE IV GLS FE IV

Dependent Variableco‐author 

weighted countitapplied countit basic      countit  countit countit

Collaboration stock 5 yrs 5 yrs 5 yrs 2 yrs 5 yrs 

Sample Full Full Full Full 10 yr BPƚƚ

Funding and Collaboration:

had some EPSRC fundingit‐1 0.079 0.100 0.184*** 0.580*** 0.243**

(0.051) (0.074) (0.054) (0.142) (0.104)

fraction of EPSRC funding with industryit‐1 0.720*** 0.952*** 0.028 0.992*** 0.994**

(0.214) (0.307) (0.209) (0.331) (0.438)

fraction of EPSRC funding with industry2it‐1 ‐1.535*** ‐1.785*** ‐0.509 ‐3.228*** ‐2.154***

(0.395) (0.575) (0.399) (0.661) (0.811)

Patent Filed:

patentit 0.029** 0.030 0.034** 0.050** 0.043*

(0.013) (0.020) (0.015) (0.020) (0.026)

patentit‐1 0.028** 0.014 0.012 0.042** 0.013

(0.013) (0.020) (0.014) (0.020) (0.024)

patentit‐2 0.026* 0.045** 0.021 0.043** 0.043*

(0.013) (0.020) (0.014) (0.021) (0.026)

Academic Rank:

senior lecturerit‐1 0.012 0.017 ‐0.004 0.053*** ‐0.016

(0.011) (0.016) (0.012) (0.016) (0.023)

readerit‐1 ‐0.003 ‐0.000 ‐0.006 0.014 ‐0.033

(0.017) (0.025) (0.018) (0.025) (0.033)

professorit‐1 ‐0.034* 0.002 ‐0.010 ‐0.015 ‐0.034

(0.021) (0.030) (0.022) (0.031) (0.039)

Constant 0.282*** 0.283*** 0.106*** 0.236*** 0.536***

(0.024) (0.036) (0.026) (0.071) (0.097)

Number of observations 28508 28508 28508 31837 16250

Number of ID 3975 3975 3975 4436 1625

R^2 (overall) 0.001 0.054 0.03 0.131 0.197

F 8.80*** 7.93*** 2.12*** 13.119*** 6.986***

Robust standard errors in parentheses, * p  < 0.10, ** p  < 0.05, *** p  < 0.01.

ǂ The dependent variable and the fraction of EPSRC funding with industry  are in logarithms. 

ƚƚ Balanced panel of 10 years.

Table 4: Impact of Industry Collaboration for other Measures of Research Output ǂ

1 2 3 4 5

Model GLS FE IV GLS FE IV GLS FE IV GLS FE IV GLS FE IV

Dependent Variable countit countit countit countit countit

Reference group (group indicator it‐1=0) high collab high collab high funding Russell group non professor

Non‐reference group (group indicator it‐1=1) low collab ƚ  low collabƚƚ low funding ƚƚƚ non Russell group  professor

Funding and Collaboration:

had some EPSRC fundingit‐1 0.236** 0.142 0.161* 0.175** 0.161*

(0.105) (0.107) (0.087) (‐0.088) (0.087)

group indicatorit‐1 * had some EPSRC fundingit‐1 ‐0.090 0.047 0.004 ‐0.1 0.259

(0.129) (0.139) (0.048) (‐0.087) (0.190)

fraction of EPSRC funding with industryit‐1 1.364*** 1.580*** 1.270*** 1.051*** 0.862**

(0.414) (0.430) (0.392) (‐0.389) (0.382)

group indicatorit‐1 * fraction of EPSRC funding with industryit‐1 ‐0.846** ‐0.509 ‐0.073 ‐0.038 ‐0.812

(0.359) (0.389) (0.291) (‐0.344) (0.612)

fraction of ESPRC funding with industry2it‐1 ‐2.687*** ‐2.723*** ‐2.177*** ‐1.386* ‐1.677**

(0.814) (0.845) (0.743) (‐0.74) (0.810)

group indicatorit‐1 * fraction of EPSRC funding with industry2it‐1 1.810** ‐0.120 ‐0.685 ‐0.676 1.511

(0.837) (0.902) (0.700) (‐0.809) (1.212)

Patent Filed:

patentit 0.040* 0.029 0.039* 0.040* 0.040*

(0.021) (0.021) (0.021) (‐0.021) (0.021)

patentit‐1 0.036* 0.031 0.034* 0.037* 0.037*

(0.021) (0.021) (0.020) (‐0.02) (0.020)

patentit‐2 0.040* 0.031 0.036* 0.038* 0.039*

(0.021) (0.022) (0.021) (‐0.021) (0.021)

Academic Rank:

senior lecturerit‐1 0.025 0.017 0.022 0.027 0.030

(0.019) (0.020) (0.019) (‐0.019) (0.019)

readerit‐1 ‐0.004 ‐0.015 ‐0.012 ‐0.002 0.008

(0.027) (0.029) (0.027) (‐0.027) (0.028)

professorit‐1 ‐0.006 ‐0.028 ‐0.024 ‐0.004 ‐0.091

(0.033) (0.034) (0.033) (‐0.032) (0.144)

Constant 0.490*** 0.573*** 0.518*** 0.451*** 0.502***

(0.040) (0.086) (0.042) (‐0.079) (0.082)

Number of observations 27667 23645 28508 28508 28508

Number of ID 3833 3150 3975 3975 3975

R^2 (overall) 0.105 0.053 0.180 0.101 0.089

F 7.963*** 8.209*** 9.637*** 8.669*** 7.923***

Robust standard errors in parentheses,  * p  < 0.10, ** p  < 0.05, 

*** p  < 0.01.

ƚƚ Based on being below the median in the percentage of average lifetime EPSRC funding with the industry. Academics with zero funding excluded.

ƚƚ Based on being below the median in the amount of EPSRC funding received in the past 5 years.

Table 5: Impact of Industry Collaboration by Groups of Academicsǂ

ǂ The dependent variable and the  fraction of EPSRC funding with industry  are in logarithms. 

ƚ Based on being below the median in the percentage of publications co‐authored with the industry. Academics with zero publications excluded.

34

Figures

+Industry 

CollaborationPublications

Ideas

Time/Attention

Resources

Constraints

+

+

-

+

+

+

−

+

Figure 1. Channels through which collaboration affects/is affected by publications

Talented &industry‐savvy

Talented &output‐driven

Talented &selective

+

-/+

-

+

+

+

+

High‐quality partner

Selectio

n 

mechanism

sMain

 channels

+

MatchingIndividual characteristics

+

35

Figure 2: Estimated publications by fraction of EPSRC funding with industry.

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.10

1.20

1.30

1.40

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Predicted number of publications

Fraction of EPSRC funding with industry

GLS FE GLS FE IV (benchmark model)

GLS FE IV lags GMM AB

GLS FE IV (no EPSRC funding)