industry or academia, basic or applied? career choices and...

MANAGEMENT SCIENCEArticles in Advance, pp. 1–21ISSN 0025-1909 (print) � ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1120.1582

© 2012 INFORMS

Industry or Academia, Basic or Applied? CareerChoices and Earnings Trajectories of Scientists

Rajshree AgarwalR. H. Smith School of Business, University of Maryland, College Park, Maryland 20742, [email protected]

Atsushi OhyamaGraduate School of Economics and Business Administration, Hokkaido University,

Sapporo, Hokkaido 060-0809, Japan, [email protected]

We extend life cycle models of human capital investments by incorporating matching theory to examinethe sorting pattern of heterogeneous scientists into different career trajectories. We link differences in

physical capital investments and complementarities between basic and applied scientists across industry andacademic settings to individual differences in scientist ability and preferences to predict an equilibrium matchingof scientists to careers and to their earnings evolution. Our empirical analysis, using the National ScienceFoundation’s Scientists and Engineers Statistical Data System database, is consistent with theoretical predictionsof (i) sorting by ability into basic versus applied science among academic scientists, but not among industryscientists; and (ii) sorting by higher taste for nonmonetary returns into academia over industry. The evolutionof an earnings profile is consistent with these sorting patterns: the earnings trajectories of basic and appliedscientists are distinct from each other in academia but are similar in industry.

Key words : human capital; matching; scientists and engineers; earnings profileHistory : Received July 19, 2010; accepted April 3, 2012, by Lee Fleming, entrepreneurship and innovation.

Published online in Articles in Advance.

1. IntroductionThe dynamism of the U.S. economy has beenattributed to advances made in academic and indus-try settings in both basic and applied domains. Theunderlying scientific labor markets can be charac-terized by two-sided matching: Hiring institutionschoose among scientists who differ in their abil-ity and preferences, and scientists—particularly asthey embark on their careers—choose among careeroptions regarding what to do—basic or appliedscience—and where to do it—academia or industry.The match of job and scientist characteristics may alsohave path-dependent outcomes and lasting implica-tions on their earnings trajectories.

However, a systematic study that accounts for sort-ing of scientists across each of the four career optionssimultaneously is lacking. Although the basic ver-sus applied nature of research is recognized as animportant characteristic in both academia and indus-try (Sauermann and Stephan 2012), scholars havelargely examined these career options two at a time.For example, studies have focused on the choicebetween basic and applied research in academic set-tings (Levin and Stephan 1991, Stuart and Ding 2006,Thursby and Thursby 2002) or in industry settings(Cohen and Levinthal 1990, Rosenberg 1990, Stern2004), or the choice between industry and academia

(Roach and Sauermann 2010). For the underlyingtheoretical mechanisms, scholars have relied on lifecycle models of human capital investments (Levinand Stephan 1991, Thursby et al. 2007) or on differ-ences in taste for science and nonpecuniary returns(Stern 2004, Roach and Sauermann 2010). These stud-ies have provided valuable insights and generatedmany empirical findings regarding outcomes of sci-entific labor markets; nonetheless, we do not have aunified theoretical model that examines the followingresearch question: How do (demand-side) differencesin complementarities between basic and applied sci-ence across industry and academic settings interactwith (supply-side) differences in ability and prefer-ences among scientists to impact their career choicesand their resultant earnings evolution?

Our study answers the above research questionby investigating sorting patterns of scientists intoalternative careers and the earnings trajectories asso-ciated with each career option. Our model buildson the premises that (i) academic and industry set-tings provide different access to physical capital forbasic and applied scientists and differ in the comple-mentarity between basic and applied scientists, and(ii) scientists differ in their ability and preferencesfor nonmonetary “taste for science.” We show thatcompetition among institutions and among scientists

1

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

Published online ahead of print October 8, 2012

Agarwal and Ohyama: Career Choices and Earnings Trajectories of Scientists2 Management Science, Articles in Advance, pp. 1–21, © 2012 INFORMS

for the most desirable partners results in a positiveassortative matching: more able scientists are matchedwith institutions that provide them more physicalcapital investments, and, when applicable, more ablebasic scientists are matched with more able appliedscientists. Further, the taste for nonmonetary returnsseparates industrial and academic scientists in equi-librium. For earnings evolution, our model predictsthat complementarities between basic and applied sci-entists in industry relative to academia results in apositive assortative matching between them, imply-ing similar earnings for basic and applied indus-trial scientists. Academic scientists’ earnings are offsetbecause of their taste for nonmonetary returns, andthe relative absence of complementarities betweenbasic and applied scientists and differences in accessto physical capital result in a divergence in theirearnings. We find support for the model implica-tions using the longitudinal Scientists and EngineersStatistical Data System (SESTAT) developed by theNational Science Foundation (NSF) for the 1995–2006period.

We contribute to Stephan’s (1996) call for bettermodeling of the labor markets in science by incor-porating matching theory (Becker 1973) into tradi-tional life cycle models of human capital investments(Becker 1962). In doing so, our study provides use-ful guidance on how possible self- or sample-selectionissues in life cycle models may impact empiricalfindings. Our model builds from the differences inthe scientific production functions in academia andindustry: basic scientists may work independentlyof applied scientists in academia, but have to workclosely with applied scientists in industry. Becauseacademic institutions make higher per capita invest-ments in basic research relative to applied, our modelgenerates some novel implications that are backedby the empirical analysis: in academia, scientists ofhigher ability sort into basic rather than appliedresearch, and initial earnings of basic scientists arelower but the slope of their earnings is higher relativeto applied scientists. In industry, by contrast, there isno such ability sorting, and the earnings trajectoriesof basic and applied scientists are similar.

Additionally, by examining all four career optionsin tandem rather than any two in isolation, our modelis able to generate other implications that extend thefindings from earlier studies. In particular, we con-tribute to the nascent literature that examines earlycareer selection of scientists because of ability andpreferences (Roach and Sauermann 2010, Stern 2004)by adding the life cycle component. We highlightthe fact that observed earnings differentials acrossthe four career options may stem from differences inhuman and physical capital investments and may notentirely be a result of a taste for science. For example,

consistent with Roach and Sauermann (2010), wefind that scientists with a higher nonmonetary “tastefor science” are more likely to choose academiathan industry, thus foregoing higher earnings. How-ever, even within academia, taking productivity andreturns from human capital investments into account,we find that earnings of academic basic scientists atlater life stages “catch up” with industry earnings.In his cross-sectional study examining industry joboffers of postdoctoral biologists at top-tier researchinstitutions, Stern (2004) found that the “preferenceeffect” overshadowed the “productivity effect,” sothat scientists seeking jobs in industry are willing toaccept lower wage offers to engage in basic research.Our longitudinal study shows that although this maybe true for the very initial years (see Figure 2 in §4.2),basic scientists in industry enjoy the same earningsas their applied counterparts for most of their life-time. We attribute this to the enhanced productivity ofindustry basic scientists since an assortative matchingallows them to access higher levels of complementaryphysical and human capital. Importantly, by examin-ing the consequences of differences in complementar-ity of basic and applied scientists across institutionalsettings, our study also has numerous managerialand policy implications. Chief among them is theneed, within universities that train young scientists,to accommodate the requirements for success of thevarious career paths that they may choose.

2. Theoretical Analysis2.1. Labor Markets for Science: Received

Literature and Stylized Features

2.1.1. Demand for Scientists Across InstitutionalSettings. In her influential literature review, Stephan(1996) describes salient aspects of the demand forscientists, which stems largely from institutions ofhigher education (academia) and for-profit businesses(industry), with government and nonprofit organiza-tions being a distinct minority. The two institutionalsettings—academia and industry—seek scientists fordifferent reasons, are governed by different normsand incentive structures, and—importantly—differin the relative emphasis on basic and applied sci-ence. Consider, for instance, the statistics reportedin Table 1 on U.S. 2003 research expenditures andemployment of scientists. The total research expen-ditures in academia and industry are the same,but the proportions spent on basic and appliedresearch are vastly different. Basic research accountsfor 74% of academic research expenditures but only17% of industry expenditures. In contrast, appliedresearch is only 26% of academic research expen-ditures but 83% in industry. The employment and

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

Agarwal and Ohyama: Career Choices and Earnings Trajectories of ScientistsManagement Science, Articles in Advance, pp. 1–21, © 2012 INFORMS 3

Table 1 Research Expenditures, Number of Scientists, and ResearchExpenditure per Scientist

Academia Industry

Basic Applied Basic Applied

Research expenditure 27,956 9,721 6,526 30,883(millions dollars)

Number of scientists 204,542 167,865 104,393 310,569Research expenditure 0.14 0.06 0.06 0.10

per scientist(millions dollars)

Sources. National Science Board (2010) (research expenditures data fromAppendix Tables 4-4, 4-5, 5-1, and 5-2) and SESTAT 2003 (labor force data,estimated using sample weights).Note. Numbers are for the year 2003 and expressed as 2000 constant dollars.

research expenditures per scientist also mirror the dif-ferential focus within each institutional setting: theacademic sector employs more and has higher expen-ditures per scientist in basic versus applied science,whereas the reverse is true for industry. Nonethe-less, more than one-third of the scientists pursue “off-diagonal” careers of applied research in academia andbasic research in industry.

Within academia, the primary focus on basic sci-ence is fueled by social norms and reward structuresthat promote nonpecuniary motives such as priorityof discovery, recognition of merit awards, and rep-utation (Merton 1973). Also, although applied sci-ence is rooted in the land grant mission of manypublic universities, the influential Bush (1945) reportto President Roosevelt resulted in a shift in focustoward basic science through increased funding andcreation of government agencies such as the NSF.Together, these forces resulted in universities plac-ing more emphasis on discoveries of scientific fun-damentals than applications of scientific discoveries(Argyres and Liebskind 1998, Goldfarb 2008, Murrayand Stern 2007). Recently and after the Bayh-Dohl act,Etzkowitz (1998) highlights the “second revolution”in academia and the increased attention to the com-mercialization of science. However, the connectivitybetween basic and applied research or research rel-evant to Pasteur’s quadrant (Stokes 1997) is largelyrealized only for certain areas of specialization(e.g., biotechnology) and/or follows a “science-push”model, where basic research is conducted prior to thepursuit of potential applications through subsequenttechnology transfer or entrepreneurship (Aghion et al.2008, Bercovitz and Feldman 2008, Stuart and Ding2006, Thursby and Thursby 2002). Tenure normsimply that in the early career years, there are fewerincentives to engage simultaneously in basic andapplied research or for basic and applied scien-tists to work together (Boardman and Ponomariov2007, Boardman and Bozeman 2007, Braxton et al.

2002, Geisler 1989). Boardman and Ponomariov (2007)and Boardman and Bozeman (2007) report that earlycareer basic and applied scientists tend to work in iso-lation of each other, not only in traditional academicdepartments, but even in “multidisciplinary, multi-purpose” or “cooperative” university research centerswhose primary responsibilities include applications ofscientific knowledge. They provide quantitative andqualitative evidence that center-affiliated basic scien-tists in early career stages (pretenure) are reluctant toinvest an effort in working with their applied counter-parts and note significant “role strain” and “shirkingof responsibilities” in areas where they need to con-sult with applied and industry partners.

Within industry, the demand for scientists stemsfrom the need to innovate (Stephan 1996), with thegoal of transforming scientific knowledge into com-mercially valuable outputs and appropriating eco-nomic value in the form of profits (e.g., Gittelman andKogut 2003, Aghion et al. 2008). Thus, the demandfor applied research is greater, and if firms engagein basic research, it is as a byproduct or coproductof their applied endeavors (Rosenberg 1990). Fun-damental breakthroughs in basic science may occurin industrial labs, given their need as a founda-tion for applied work (Hounshell and Smith 1988,Rosenberg 1990, Stokes 1997); alternatively, firms mayengage in basic science to create absorptive capac-ity (Cohen and Levinthal 1990) and enhance the pro-ductivity of their R&D efforts (Griliches 1986). Anec-dotally, Hounshell and Smith (1988) highlight howbasic science at DuPont served to search for newfields of chemistry, provided foundations for scientificinvestigation of DuPont’s existing technology, andshowcased the company’s technical competence andcapabilities. Importantly, these authors underscorethe need for basic and applied scientists workingtogether, best exemplified by the following state-ment by DuPont’s basic research department direc-tor Charles Stine: “Perhaps the most important ideawhich I have attempted to keep before my assistantsin this work is the necessity for continuous intimatecontact with the various departments and subsidiarieswhich we have been attempting to serve” (Hounshelland Smith 1988, p. 137). Thus, relative to academia,where basic and applied scientists can work in paral-lel or independently of each other, sequencing theseactivities over time, or for different purposes, there isa more direct link and greater synergies between basicand applied scientists in industry.1

1 These distinctions in industry/academia highlight the extreme dif-ferences that are salient at the time a scientist embarks on hercareer. Of course, a continuum of possibilities may occur, particu-larly in later career stages. As Stephan (1996) notes, academic sci-entists often engage in privatization of knowledge, and industrialscientists often disclose their knowledge voluntarily.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


2.1.2. Supply of Scientists and Heterogeneityof Ability and Preferences. On the supply side,research ability and preferences have been high-lighted as distinctive characteristics of scientists thataffect occupational and activity choices and compen-sation structure (Dasgupta and David 1994, Stephan1996, Stern 2004). People differ in their abilities tolearn extant knowledge and develop new innova-tions, either because of innate differences in intelli-gence or prior investments in knowledge. Also, scien-tists purposefully raise the level of their human cap-ital and therefore their earnings capacity over timeby conducting scientific research (Becker 1962, Ben-Porath 1967, Levin and Stephan 1991, Thursby et al.2007). The supply-side dynamics of labor markets inscience accordingly are often modeled in a life cycleframework wherein human capital investments areundertaken to maximize lifetime utility that includesboth monetary and nonmonetary benefits (Levin andStephan 1991). The inclusion of nonmonetary bene-fits is consistent with the characterization of the sci-entist’s preferences or “taste for science” indepen-dent of ability, which may vary based on the questfor basic research and the desire to apply scientificprinciples for economic and technological develop-ment (Levin and Stephan 1991, Roach and Sauer-mann 2010, Stephan 1996, Stern 2004). Such prefer-ences reflect attitudes toward open science, freedomof research topics, and intellectual challenge and areinformed in the socialization process when receivingscientific education (Stephan 1996). Roach and Sauer-mann (2010) show that scientists with a higher tastefor science prefer careers in academia over industry.Stern (2004) discusses the strong positive correlationbetween ability and preferences observed among sci-entists. Although higher-ability scientists commandhigher compensation, they may also have a greatertaste for science, resulting in their willingness toaccept a lower wage as a compensating differential.

2.1.3. Implications for Modeling Early CareerChoices. In concluding her literature review, Stephan(1996) noted that extant human capital and life cyclemodels do not capture the complexities of the pro-duction of scientific knowledge. In part, this may bebecause most models do not study the supply anddemand side simultaneously. Extending human capi-tal and life cycle models with matching models mayresult in a more appropriate characterization of scien-tific labor markets and shed light on factors that influ-ence early career choices. Matching models (Becker1973, Roth and Sotomayor 1992, Sattinger 1993) areparticularly suitable for the study of scientific labormarkets because they are better able to accommodatemarket outcomes in exchanges of heterogeneous andindivisible “goods” (Mindruta 2012).

In developing a model of scientists’ early careerchoices, we incorporate the above salient demand

and supply characteristics of scientific labor markets.On the supply side, we assume heterogeneity amongscientists in ability and preferences for nonpecuniaryreturns. On the demand side, the career options rep-resent differences in the physical and human capitalthat complement a scientist’s human capital. Withina multifactor scientific production function context,we define “complementarity” as the positive effect onone input’s marginal productivity due to a marginalincrease in the other factor. We focus on two typesof complementarities. The first is between a scien-tist and the physical capital provided by the institu-tion to her. We assume, based on Table 1, that thebasic scientist has greater access to physical capitalthan the applied scientist within academia, and thereverse holds in industry. The second type of com-plementarity is between basic and applied scientists.We employ a simplifying assumption that the focalscientist is either a basic or an applied scientist but notboth, thus ruling out the potential that an individualscientist may undertake basic and applied scientifictasks simultaneously (Mansfield 1995). This assump-tion is consistent with early career realities that sci-entists have to focus on one primary activity initially.We later discuss how the model implications maychange when allowing for career switches or tran-sitions to Pasteur’s quadrant (Stokes 1997) by moresenior scientists. At least in the initial (pretenure)years of their career, basic academic scientists areless likely to work with their applied counterparts(Boardman and Ponomariov 2007, Boardman andBozeman 2007). Within industry, firms that engagein basic research expect basic and applied scientiststo work with each other on issues that have poten-tial applications of knowledge for the firm (Hounshelland Smith 1988, Cohen and Levinthal 1990). Accord-ingly, we assume that basic and applied scientistscomplement each other in industry knowledge pro-duction functions, but not in academia.

2.2. A Brief, NonmathematicalOverview of the Model

This section verbally describes the formal modeland propositions in §§2.3 and 2.4 to explicate theunderlying logic and intuition. The building blocks,mechanisms, and resultant propositions are also sum-marized in Figure 1. A scientist’s career is definedby a combination of the type of research she con-ducts (basic or applied) and the type of institutionshe works in (academia or industry), resulting infour career options: (i) basic scientist in academia,(ii) applied scientist in academia, (iii) basic scientist inindustry, and (iv) applied scientist in industry. A sci-entist supplies her human capital, and the institutionshe works for supplies the complementary physicaland other human capital necessary to conduct scien-tific research.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Figure 1 Verbal Description of Model Assumptions, Mechanisms, and Predictions

Underlying Assumptions Mechanisms/Forces in Model Model Implications/Predictions

Main individualcharacteristics

C1: Scientists differ in ability

C2: Scientists differ in theirtaste for nonpecuniaryreturns

Main institutioncharacteristics

C3: Strong complementaritybetween basic and appliedscientists in industry but notin academia

C4: Basic scientists havegreater access to physicalcapital relative to appliedscientists in academia;the opposite holds inindustry

Main forces affecting careerchoices and earnings

trajectories

F1: Superior access to physicalcapital attracts more ablescientists

F2: Complementarity betweenbasic and applied scientists inindustry makes the effectiveaccess to physical capitalthe same for each type ofscientists

F3: Complementarity and positiveassortative matching betweenbasic and applied scientists inindustry results in synergy effectson earnings in equilibrium

F4: Academic researchenvironments offer nonmonetaryreturns

F5: Superior access to physicalcapital makes human capitalinvestments more productive

Empirical predictions related to career choices

Proposition 1 for academia: Given C1, C3, and C4, F1 implies

In academia, the average research ability of basic scientists is higherthan that of applied scientists.

Proposition 1 for industry: Given C1, C3, and C4, F2 implies

In industry, the average research ability does not depend on whetherresearch type is basic or applied.

Proposition 2: Given C2 and C3, F3 and F4 imply

If nonpecuniary returns are sufficiently large and physical capitalavailability is relatively modest, those scientists who valuenonpecuniary returns are more likely to choose academiaover industry.

Empirical predictions related to earnings trajectories

Proposition 3: Given C2 and C3, F3 and F4 imply

The average earnings profile of academic scientists is locatedbelow those of industrial scientists when the synergy effectsin industrial research are sufficiently large.

Proposition 4: Given C1, C3, and C4, F1 and F5 imply

The average initial salary of academic basic scientists is lower thanthat of academic applied scientists.

Proposition 5 for academia: Given C1, C3, and C4, F1 and F5 imply

The expected value of the slope of the earnings profile is higher forbasic scientists than for applied scientists in academia.

Proposition 5 for industry: Given C1, C3, and C4, F2 implies

The expected value of the slope of the earnings profile is similar forbasic and applied scientists in industry.

Scientists choose a career option to maximize life-time utility based on pecuniary and nonpecuniaryreturns (Roach and Sauermann 2010, Stern 2004)and purposefully invest to raise their human capi-tal and thus earnings over time (Becker 1962, Levinand Stephan 1991, Thursby et al. 2007). Similarly,institutions choose scientists to maximize scientificknowledge production, based on their productivity.As in matching models (Becker 1973), the “match”between a particular scientist and a particular insti-tution is determined as scientists compete with eachother to obtain a position that maximizes lifetime util-ity, and institutions compete with each other to attracthighly productive scientists. The sorting of scientistsis dependent on the identity of both the scientist andthe institution. Because the productivity of humancapital investments also depends on the complemen-tary factors available through the institution, moreable scientists and institutional settings with access togreater complementary human and physical capitalare matched.

Equilibrium outcomes are driven by differences incomplementarity across institutional settings and pos-itive assortative matching. By a positive assortativematching, we mean a positive association betweenpairs of scientists or scientists and institutions withrespect to research ability and level of physical capi-tal (Becker 1973). These institutional differences imply

that basic academic scientists have access to morephysical capital than applied academic scientists,whereas the opposite is true in industry. However, inindustry, complementarity between basic and appliedscientists makes the effective access to physical capi-tal the same for each scientist type. Further, academiaprovides greater nonpecuniary returns than indus-try. In academia, low complementarity between earlycareer basic and applied scientists, combined withrelatively lower levels of physical capital in appliedscience, results in more able scientists sorting intobasic rather than applied science (Proposition 1 foracademia). In industry, complementarities encouragebasic and applied scientists with similar ability towork for the same firm, resulting in a positive assor-tative matching between basic and applied scientistsas well as between scientists and firms. Thus, there isno sorting based on ability across basic and applieddomains in industry (Proposition 1 for industry).Further, scientists with higher nonpecuniary prefer-ences are matched to academia relative to industry(Proposition 2).

The above sorting results in scientists’ earningsevolution differently within each career option. Dif-ferences in research ability and preferences fornonpecuniary returns are associated with the evo-lution of average earnings for each career path.Given sorting based on nonpecuniary returns and

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


complementarity among basic and applied scientistsin industry, which generates additional monetarygains from synergies realized from positive assorta-tive matching, the earnings profile of academic scien-tists is lower than industrial scientists (Proposition 3).Further, within academia, basic scientists invest moretime initially in building up their human capital thando applied scientists, because access to higher lev-els of physical capital makes their marginal benefitfrom human capital investments higher. As a result,initial earnings of basic academic scientists are lowerthan those of applied academic scientists (Proposi-tion 4). Finally, positive assortative matching amongbasic and applied industrial scientists results in a sim-ilar evolution of their earnings profiles. In contrast,the relative lack of complementarity leads to diver-gent paths of earnings for basic and applied scientistsin academia, with a steeper slope for basic than forapplied academic scientists (Proposition 5).

2.3. Formal Theoretical Development

2.3.1. Key Assumptions. The formal model isunderpinned by the following assumptions: (i) scien-tists are perfectly rational; (ii) there is no asymmetricinformation and no uncertainty; (iii) scientists do notchange their career over their life cycle; (iv) withinacademia, basic scientists have greater access to phys-ical capital than applied scientists; (v) within indus-try, there are two types of firms—type A firms hireapplied scientists only and type B firms hire bothapplied and basic scientists; (vi) a taste for nonmone-tary returns enters the utility function additively andexogenously; and (vii) academia offers nonpecuniarybenefits, but industry does not. Some model impli-cations hinge on these assumptions for simplificationand tractability of the theoretical analysis. In §5, wediscuss how the implications change when certainassumptions are relaxed.

2.3.2. Key Properties of Scientific KnowledgeProduction Function. We characterize the scientist’sresearch activities to set the stage for analyzing theinteraction of market forces and scientists’ prefer-ences leading to an equilibrium allocation of scientistsinto the alternative career paths. Consider a modelin which scientists make a human capital investmentto increase their knowledge base, and their accu-mulated knowledge is used as an input to producefinal research output. Formally, scientist i’s knowl-edge base evolves as

Hit+1 =Hit +Rit1 (1)

where Hit is the level of scientist i’s knowledge baseat time t, and Rit is knowledge added to the stock ofher knowledge base. To keep our exposition simple,we assume no depreciation of knowledge base.

Scientist i’s knowledge production function, Rit , inEquation (1) depends not only on how much timeshe devotes to human capital investments, but also onwhere she works (i.e., academia or industry) and whattype of research she conducts (i.e., basic or applied).Formally, it takes the form of

Rit = �k6�i4litHit5�7�1 6�j4ljtC jt

5�7�21 (2)

where �i is the ability parameter of scientist i, and lit isi’s time spent on building human capital. For scientisti in basic (applied) research, Cjt is the complemen-tary human capital provided by scientist j in applied(basic) research. Parameters �1 and �2 are either 0 or1 and control the degree of complementarity betweenbasic and applied scientists. Parameter � is exoge-nously given and takes some value between 0 and1/2. The amount of physical capital available to thescientist is denoted by �k, which is explained in detailbelow.

To model the knowledge production function inacademia, we set �1 equal to 1 and �2 to 0 for anacademic setting, which allows for the lack of comple-mentarity between basic and applied scientists. Thus,for scientist i conducting type k research in an aca-demic setting, the knowledge production function is

Rit = �k�i4litHit5�0 (3)

For industry, we model two types of firms: Type Afirms conduct only applied research and have theknowledge production function as given by (3).Type B firms conduct both applied and basic research.For scientist i conducting type k research in type Bfirms, the knowledge production function is

Rit = �k�i4litHit5��j4ljtC jt

5�0 (4)

Implications of relaxing the above assumptions,which allow us to understand essential forces leadingto an equilibrium-matching pattern but at the cost ofsome loss of generality, are discussed in §2.4.

Scientists are differentiated by ability � and a tastefor nonpecuniary returns �. We assume that � is dis-tributed according to a Pareto distribution with a den-sity function

f 4�5= a�−4a+15 for � > 11 (5)

where a≥ 2 to ensure existence of at least the secondmoment. The variable � takes either 0 or 1, indicatingthat a scientist has a taste for nonpecuniary returns if� equals 1. Because the taste for nonmonetary returnsis better accommodated in academia rather than inindustry (Sauermann and Stephan 2012), we incor-porate this factor in the model by assuming thatscientists with a taste for nonpecuniary returns gain

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


additional nonmonetary benefits x when they workin academia. The variables � and � are independentlydistributed.

In addition to human capital, physical capital isan important input of research production. Becauseacademia places relatively more emphasis on basicthan on applied science and the investments in phys-ical capital are commensurate with these differences(see Table 1), we incorporate this in our model byusing the concept of the first-order stochastic domi-nance. Physical capital per scientist, �k, is assumed tobe Pareto distributed with a density function

g4�k5= bcbk�−4b+15k for �k > ck (6)

with

b ≥ 21 cBA > cAA > 0 and cAI ≥ cBI > 01 (7)

where subscripts BA, AA, BI, and AI indicate“basic research in academia,” “applied research inacademia,” “basic research in industry,” and “appliedresearch in industry,” respectively. An essential prop-erty of the specification by (6) and (7) is the first-orderstochastic dominance relation. Under this specifica-tion, which makes the model analytically tractable,the fraction of basic academic scientists who enjoyaccess to large amounts of physical capital is greaterthan the fraction of applied academic scientists.

2.3.3. Scientists’ Predilections over Career Paths.We first derive a scientist’s optimal path of humancapital investment for each career choice and resul-tant lifetime earnings. Scientists seek to maximize thediscounted present value of their lifetime earningsby allocating their time between building up humancapital, lit , and converting accumulated human capi-tal into observable knowledge outputs (e.g., patents,publications, new products and processes, etc.) forearnings, 1 − lit . We assume that scientists are notdirectly compensated for the time invested in build-ing human capital but are paid for the conversion ofhuman capital into new knowledge (Ben-Porath 1967,Levin and Stephan 1991, Thursby et al. 2007). Becausescientists are paid according to the current level oftheir human capital (Thursby et al. 2007), their currentearnings, Yit , are

Yit =wk41 − lit5Hit1 (8)

where wk is the rental rate of human capital whenscientist i chooses career path k.2

2 Scientists’ pay is assumed to be linear to the stock of human cap-ital to avoid computational complexity, but nonlinear forms canbe incorporated to reflect that human capital is an input for finalresearch outcomes.

Time flows discretely and T is the terminal period.In her dynamic optimization problem, a scientist opti-mally divides her time between building human cap-ital, lt , and converting human capital to earn, 1 − lit ,so as to maximize her lifetime earnings,

∑Tt=0 �

tYit ,where � is a discount factor. The Bellman equationfor this maximization problem is given by

Vt4Hit5= maxlit

6wk41 − lit5Hit +�Vt+14Hit+157 (9)

subject to (1) and (3) if a scientist works in academiaor in a type A firm and to (1) and (4) if she worksin a type B firm. By using the terminal conditionVT+14HT+15 = 0 and solving (9) recursively, we obtainthe optimal path of lit . All the details of the deriva-tion in this section are placed in the online appendix,which is available from the authors upon request.Given the evolution of the complementary humancapital, the evolution of scientist i’s human capi-tal and earnings are pinned down by Equations (1)and (8), respectively. These are given by

Yt4Hit5

=wk

[

Hit − 4�k�i�Ajt51/41−�5

(

�−�T−t+1

1 −�

)1/41−�5]

(10)

and

Hit+1 =Hit+4�k�iAjt51/41−�5��/41−�5

(

�−�T−t+1

1−�

)�/41−�5

1

(11)where Ajt = �j4ljtC jt

5� if she works in a type B firmand Ajt = 1 otherwise.

Scientists are assumed to be risk neutral and theirlifetime utility depends on both monetary and non-monetary rewards. Formally, for academic and type Afirm scientist i conducting type k research, this is

U4�i1�k5=wk64�i�k51/41−�5�A +�7+ �ixi1 (12)

where �A ≡ �1/41−�5441 −�5/�5∑T−1

s=1 44�−�T−s+15/41 − �551/41−�5 and � ≡ 441 −�T 5/41 −�55Hi1. Forindustrial scientists in type A firms, �i is set to zero.The first term in (12) is the optimal lifetime earningsof scientist i when she conducts type k research inacademia. If scientist i conducts type k research intype B firms, her lifetime utility is

U4�i1�k5=wk64�i�k51/41−�5�I 4�j5+�71 (13)

where �I 4�j5≡�1/41−�5441−�5/�5∑T−1

s=1 Ajs44�−�T−s+15/41 −�551/41−�5.

Scientists have perfect foresight regarding theirstream of future earnings when they choose a careerpath because there is no uncertainty about humancapital investment. Therefore, scientists use optimal

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


lifetime earnings to anticipate monetary rewards gen-erated from each career choice, and Equations (12)and (13) completely characterize scientist i’s predilec-tions over the four alternative careers. Furthermore,the absence of uncertainty implies that scientists donot change their career path in the middle of theircareer (We discuss relaxation of this assumption §5.1).Scientist i’s predilections over the four alternativecareer paths depend on her ability, �i, a taste for non-pecuniary returns, �i as well as on what type researchshe conducts and where, �k. Most importantly, otherthings being equal, a scientist’s predilection increaseswith parameter �k of research productivity associatedwith where she works and what type of research sheis conducting.

2.3.4. Institutions’ Predilections over Scientists.We now describe predilections of institutions (univer-sity in academia and firm in industry) when they hirescientists. An institution’s net gain at time t from hir-ing scientist i in terms of monetary value is given by

�t = p641 − lit5Hit7�

−wk41 − lit5Hit0

To ease the computational burden, we assume that� = 1 so that cumulative research outputs are trans-formed linearly into commercially or academicallyvaluable outputs. Under this assumption, a net gainfor a university or a type A firm from hiring scientisti throughout her entire career is

ç=

T∑

t=1

�t−1�t = 4p−wk564�i�k51/41−�5�A +�7 (14)

and the corresponding gain for a type B firm is

ç= 4p−wk564�i�k51/41−�5�I 4�j5+�70 (15)

Thus, given that a firm or university hires scientiststo maximize scientific knowledge output, both typesof employers have a predilection for hiring scientistswith high ability �.

2.4. Implications of the Theoretical Model

2.4.1. An Equilibrium Sorting Pattern. Based onstable matching (Becker 1973, Roth and Sotomayor1992, Sattinger 1993), we examine an equilibrium sort-ing pattern of scientists. Given limited availability ofspecific positions, scientists compete with each otherfor a desirable career option. Similarly, the scarcityof able scientists leads research institutions to com-pete with each other to attract desirable scientists. Forthe sake of simplicity, we assume that a firm and auniversity both hire a pair of scientists. A matchingoutcome is a set of disjoint triplets (scientist i, insti-tution j , type of research k5 that indicates that sci-entist i conducts research type k (applied or basic)

in institution j (academia or industry). A match isstable if there is no other triplet that makes botha scientist and an institution better off (Roth andSotomayor 1992). In the preceding section, we showedthat predilections of both scientists and research insti-tutions strictly increase with research productivity.In particular, the complementarity in knowledge pro-duction function (3) or (4) results in scientists’ andresearch institutions’ predilections (12)–(15). Standardmatching theory thus implies that a stable match isa positive sorting with respect to these characteris-tics. In both academia and industry, complementaritybetween a scientist’s ability � and a university’s orfirm’s physical capital � encourages highly able sci-entists to be matched with institutions that providethem greater access to physical capital.

Lemma 1. In both academia and industry, a positivesorting with respect to a scientist’s ability � and an insti-tution’s physical capital � takes place in equilibrium.

All the formal proofs of the lemma and propo-sitions in this section are in the online appendix.Lemma 1 has implications for differences in the aver-age quality of scientists across different career cate-gories. To see this, Lemma 1 implies that demand andsupply of scientific ability in academia satisfy

cbAA

∫ �

�bx−4b+15 dx+ cbBA

�∫

�

bx−4b+15 dx = 2∫ �

�ax−4a+15 dx1

which can be simplified to

�4�5= dA�b/a1 (16)

where dA ≡ 42/4cAA + cBA551/a < 1.

In conjunction with (6) and (7), which state first-order stochastic dominance of the distribution ofphysical capital available to basic scientists relative toapplied scientists in academia, (16) implies that theaverage value of � for academic basic scientists ishigher than that for academic applied scientists: moreable scientists choose basic science given complemen-tarity between human and physical capital.

In addition to the complementarity between sci-entists and physical capital, there is complementar-ity between basic and applied scientists in type Bfirms. Our model allows scientists to choose a sideof the match between basic and applied scientists.As Kremer and Maskin (1996) extensively analyze,complementarity does not suffice for a positive assor-tative matching between agents when agents canchoose a side. In particular, they point out that a pos-itive assortative matching between agents may notoccur when the amount of outputs from joint produc-tion depends on who in a matched pair does which

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


task. Although complementarity encourages the for-mation of a match of the same traits, this asymmet-ric force discourages it and creates a gap in traits ofmatched individuals. In our knowledge productionfunction (4), agents’ productivities �i and �j enter ina perfectly symmetric way so that, for a given levelof physical capital, the asymmetric force is totallyabsent. In other words, we assume that the totalamount of knowledge created is unchanged if roles ofbasic and applied scientists are exchanged. Further-more, a scientist has access to her partner’s physicalcapital indirectly through a matching in equilibrium.This implies that effective access to physical capitaldoes not depend on whether she conducts basic orapplied research. Therefore, a basic (applied) scientistwith a given level of research ability is matched withan applied (basic) scientist with that same level ofresearch ability. Under this assumption, the followingproposition holds:

Proposition 1. The average research ability of basicscientists in academia is higher than that of applied scien-tists in academia. In industry, the average research abilitydoes not depend on whether the research type is basic orapplied.

A scientist’s choice between academia and indus-try depends on nonmonetary returns, x, as well asgains from the complementarity. Because of Lemma 1,the difference in lifetime utility between working inindustry and academia is given by

ã= 4�− 15wz�4a+b5/4a41−�55− �x1 (17)

where � ≡ 41/254a�−41−2�541−�55/4a41−2�541−�55��/441−2�541−�55 ·

�4b+a�5/4a41−2�541−�5544cBI + cAI 5/4cBA + cAA551/41−�5, z ≡

d1/41−�5A �A and wk =w for any k.The first term is the monetary gain from the com-

plementarity from working for type B firms; the sec-ond term is the nonmonetary return to working inacademia. For some scientists, � can be less than 1,but � increases with �k. In other words, the synergyfrom positive assortative matching between basic andapplied research in type B firms is realized for pairsof highly able scientists and institutions, but such asynergy is negligible or absent if the level of humancapital of scientists and of physical capital of an insti-tution is low. In our model, in addition to the effectof taste for science, complementarities between basicand applied scientists also create lifetime earnings dif-ferentials between academic and industrial scientists.

Proposition 2. If nonpecuniary returns are suffi-ciently large and physical capital availability is relativelymodest, those scientists who value nonpecuniary returnsare more likely to choose academia over industry.

2.4.2. Earnings Evolution. This section character-izes the evolution of the earnings profile along withan equilibrium path. Equations (10) and (16) imply

that the earnings profile of scientists in academia andtype A industrial firms evolves in equilibrium accord-ing to

Yt =wk6Ht −�b/4a41−�55k d

1/41−�5A �t71 (18)

and the earnings profile of industrial scientists intype B firms evolves in equilibrium according to

Yt =wk6Ht −��b/4a41−�55k d

1/41−�5A �t71 (19)

where �t = ��/41−�544�−�T−t+15/41 −�55�/41−�5.

The above two equations, coupled with (17), sug-gest several implications for differences in earn-ings profiles among the four career options. First,we examine differentials across institutional settings.As can be seen from (17), beyond the sorting of scien-tists based on nonpecuniary returns, the model pre-dicts that the synergy effect in industry shifts theearnings profile of industrial scientists up. The mag-nitude of the synergy effects is positively correlatedwith the quality of the pool of scientists. In otherwords, the synergy effect is more likely to gener-ate monetary gains when the distribution of abil-ity has a fatter right tail. Sufficiently large synergyeffects relative to nonpecuniary returns create a posi-tive earnings differential between industrial and aca-demic scientists at any given point in time.

Proposition 3. The average earnings profile of aca-demic scientists is located below those of industrial scien-tists when the synergy effects in industrial research aresufficiently large.

Next we turn to expected differentials betweenbasic and applied scientists within each institutionalsetting. Within academia, basic scientists invest morein building up their human capital than do appliedscientists, because access to a higher level of physi-cal capital makes their marginal benefit from humancapital investments higher. Because the initial level ofhuman capital is assumed to be the same for all sci-entists, the average initial earnings of academic basicscientists are lower than those of academic appliedscientists.

Proposition 4. The average initial salary of academicbasic scientists is lower than that of academic appliedscientists.

The positive sorting also affects the earningsgrowth of academic scientists. Under positive sorting,more able scientists are on average drawn to basic sci-ence within academia. For a given level of physicalcapital, therefore, basic scientists are more productivein human capital investment than applied scientists.Additionally, on average, basic scientists have accessto a higher level of physical capital than applied sci-entists. Both factors lead basic scientists in academiato invest more in human capital at early stages of their

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


life cycle, making the earnings profile of basic scien-tists in academia steeper than that of applied scientistsin academia.

In contrast, a different mechanism is at work inindustry. For type B firms, basic and applied scien-tists are paired in research activities to exploit possiblesynergy effects (see Equation (17)). Thus, the synergyeffect shifts up the earnings profile of industrial scien-tists in type B firms, relative to the earnings profile ofapplied scientists in type A firms. This force works todirect more able scientists toward type B firms. How-ever, inequality (7) implies that applied scientists intype A firms have more favorable access to physicalcapital. This is because the effective physical capitalfor a scientist in type B firms is the average of physi-cal capital associated with a pair of basic and appliedscientists. This force shifts up the average earningsprofiles of type A industrial scientists. In equilibrium,these two opposite forces cancel out to make scien-tists with a given ability indifferent between choos-ing these two types of firms. The earnings profile ofbasic scientists in type B firms coincides with thatof applied scientists in type B firms, but may bedifferent from the earnings profile of applied scien-tists in type A firms. The average earnings profile ofindustrial applied scientists is a weighted average ofthe earnings profiles of applied scientists in types Aand B firms and is pinned down by the distributionof types A and B firms. The average earnings pro-files of basic and applied scientists in industry candiffer from one another, but the difference is smallerthan the corresponding difference in academia. In anextreme case where there are no type A firms, theearnings profile of basic scientists is the same as thatof applied scientists in industry. Accordingly,

Proposition 5. The expected value of the slope of theearnings profile is (a) higher for basic scientists than forapplied scientists in academia and (b) similar for basic andapplied scientists in industry.

3. Data and MethodologyThe empirical analysis uses the NSF restricted filesof the Survey of Doctorate Recipients (SDR) for theyears 1995, 1997, 1999, 2001, 2003, and 2006. SDR,a part of SESTAT, contains information about theemployment, education, and demographic character-istics of scientists and engineers in the United States.3

In using the SDR database, we focus on doctoratedegree recipients from U.S. institutions, given that adoctorate is usually a necessary qualification for bothbasic and academic research occupations in scienceand engineering fields. Further, the complementary

3 For the details of SESTAT, see http://www.nsf.gov/statistics/sestat/.

relationship in knowledge production between basicand applied scientists in industry is much more likelyto exist for highly specialized scientists such as doc-toral degree holders than for graduates who holdonly a bachelor’s or master’s degree. The inclusion ofbachelor’s and master’s degree holders from the inte-grated SESTAT database does not change our empir-ical results significantly, but it blurs the main pointsof our analyses. Unless otherwise noted, we focus ourempirical analysis on scientists and engineers whoare employed in the four career options of interest,are aged 65 or younger, report nonzero annualizedbasic salary, and work full time (working weekly forat least 30 hours and annually for at least 48 weeks).4

Postdocs are included in our empirical analysis, butdoctoral recipients working in teaching and consult-ing positions are excluded. See Table I in the onlineappendix for a description of criteria based on whichindividuals are included/excluded from the sample.As seen in Table 2, the total number of individuals inthe sample from 1995 to 2006 is 33,776. These longitu-dinal data are the main data for our empirical exam-inations, but repeated cross-sectional and panel dataare also used to ensure that our results are robust tocohort effects.

3.1. Variable Definitions and Descriptive Statistics

3.1.1. Career Options. Our career choice variableis obtained using the responses on two questions inthe SDRs. The first question relates to the type of prin-cipal employment, and the second relates to the typeof job activity on which the majority of time is spent.For the first question, respondents who reported thattheir employer was a four-year college or university,medical school, or university-affiliated research insti-tute are identified as working in academia, whereasrespondents who reported that their employer wasprivate for profit are identified as working in industry.For the second question, respondents who reportedbasic (applied) research as their primary job activityare identified as such. Combining these two piecesof information permits the sorting of scientists in thedata into each of the four career options. We recognizethe lack of consensus on the definition and measure-ment of the “basic” or “applied” nature of researchacross various prior studies (Murray and Stern 2007,Sauermann and Stephan 2012) and that research maybe within a continuum. The reliance on questions that

4 Academic contracts may consist of either 12-month appointmentsor nine-month appended with summer contracts ranging from oneto three months. Seventy percent of all academics in the databasereport working at least 48 weeks (approximately 11 months). Theresults are robust to analysis where we included the additional 20%of academic scientists that work between 36 and 47 weeks (corre-sponding to a minimum of a nine-month contract).

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 2 Descriptive Statistics

Academia Industry

Basic Applied Basic Applied All

DemographicsAge 41044 43019 40052 42005 42004Gender (male) 0067 0067 0075 0080 0072White 0052 0052 0050 0051 0051Married 0075 0076 0074 0080 0077U.S. citizen 0085 0086 0080 0085 0085

EarningsReal annual earnings 55,285 59,236 74,365 78,632 65,288

Scientist abilityYears to bachelor degree 4007 4040 4035 4011 4017School ranking 3035 3012 3037 3021 3025Grant recipient 0059 0042 0045 0041 0047Father’s education— 0053 0044 0053 0045 00484-year college or higherMother’s education— 0037 0029 0034 0028 00314-year college or higher

Job importanceSalary 3039 3038 3050 3045 3043Benefits 3052 3049 3050 3050 3050Job security 3051 3051 3040 3042 3046Job location 3047 3043 3040 3040 3043Opportunity for 3050 3050 3043 3041 3047advancementIntellectual challenge 3086 3087 3080 3077 3083Level of responsibility 3044 3045 3040 3038 3042Degree of independence 3076 3075 3063 3060 3070Contribution to society 3056 3053 3031 3032 3045

Number of observations 12,370 8,084 1,194 12,128 33,776

Source. Authors’ estimation using restricted use SDR data for 1995–2006.Notes. A subsample from 1995, 1997, and 1999 SDRs is used for statisticson scientist ability. For job importance, 2001 and 2003 SDRs are used toproduce descriptive statistics.

ask scientists to self-report their primary job activityas one still permits them to have aspects of the otherwhile nonetheless categorizing themselves in a pri-mary activity, and alleviates some of the problems ofthe other measures, as described by Sauermann andStephan (2012).

3.1.2. Earnings Profile. Survey respondents re-port their annualized salary, excluding bonuses, over-time or other additional compensation. We deflatethese annualized earnings by the consumer priceindex (base year is 1995). As seen in Table 2, aver-age earnings (averaged over individuals and time) ofapplied scientists in industry are the highest amongthe four groups, and the average earnings of basic sci-entists in academia are the lowest.5 Notably, there isan approximately $23,000 annual earnings differentialbetween these two groups.

5 Given the 48 weeks threshold for full-time workers noted above,these salary figures include at least two months of “summer” com-pensation for academics on nine-month contracts.

3.1.3. Scientist Characteristics. For characteristicsof scientists, key variables in our theoretical modelrelate to their research ability �, a taste for nonpecu-niary returns �, and accumulation of human capitalover time. Because there are no universal empiricalmeasures for these unobservables, we utilize severalpieces of information from the survey to proxy forthese variables. To proxy for differences in researchability (time invariant �), we rely on informationrelated to (i) time to complete first baccalaureatedegree, (ii) Ph.D. program ranking by field of science,(iii) whether a scientist received a grant during herdoctoral program, and (iv) parental educational lev-els.6 We obtain data on Ph.D. program ranking fromthe National Research Council’s evaluation of Ph.D.program quality (Golderberger et al. 1995), and otherability measures are within SDR. These variables rep-resent variations in quality and capture the scientist’sability prior to entering in the labor market; thus,they should correlate highly with their research abil-ity. To gauge preferences and determine proxies fora taste for nonpecuniary returns �, we utilize sur-vey responses to questions asking scientists to ratethe relative importance they placed on the followingjob characteristics: (i) opportunity for advancement,(ii) benefits, (iii) intellectual challenge, (iv) indepen-dence, (v) location, (vi) responsibility, (vii) salary, (viii)security, and (ix) contribution to society. The answerswere coded on a four-point scale, one being notimportant at all to four being very important.

We use a variable for labor market experienceto estimate how the earnings profile in each careerevolves over time (Propositions 3–5). This variableis measured as age minus years of schooling minussix years, and we include both linear and quadraticterms in the earnings equation. Instead of labor expe-rience, age is included as a control in the cross-sectional regressions testing scientists’ career choices(Propositions 1 and 2). Additionally, we use severaldemographic control variables such as dummies forgender, race, marital status, and U.S. citizenship, withone denoting male, white, married, and U.S. citizen,respectively. Table 2 provides descriptive statistics forthe data used in the analysis.

3.2. MethodologyPropositions 1 and 2 (career choices) are tested usinga probit model and are robust to logit, multino-mial logit, or linear probability specifications. For

6 The relationship between parents’ and children’s ability, partic-ularly as measured by educational levels, has been well docu-mented in prior studies (cf. review by Haveman and Wolfe 1995),and attributed to both biological (i.e., genetic) and environmental(e.g., home investments) factors (Leibowitz 1974, Sacerdote 2000).

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Proposition 1, the following model is tested separatelyfor academia/industry:

Pr4Dji =15=ê4�0j +�1j abilitymeasuresji+�2jXji51 (20)

where the dummy dependent variable takes the valueof 1 if a scientist i conducts basic research and 0 ifapplied research within each institutional setting j .We use information from the 1995, 1997, and 1999SDR files, given lack of relevant information for theother years. In addition to the reported empirical mea-sures of an individual’s research ability, we includedthe demographic variables listed above and the sci-entist’s educational field as controls, although thesecoefficients are not reported. For Proposition 2, weuse the 2001 and 2003 SDR files, because the jobimportance questions were only asked in these twosurvey years. The dependent variable in (20) is modi-fied to indicate whether a scientist works in academiaor industry, and ability measures are replaced byjob importance variables. The analysis reported inTables 3 and 4 for Proposition 1 and 2 is conductedby pooling the cross-section data across the relevantsurvey years, with one observation per individual toensure independence of observations.

Propositions 3, 4, and 5 related to earnings profileof scientists are tested using all SDR files and several

Table 3 Test for Proposition 1 (Choice of Basic/Applied ResearchWithin Each Institutional Setting)

Academia Industry

Years to completion of first bachelor 00009∗ −00006∗∗

4000055 4000035Grant 0013∗∗∗ 0001

400025 400015Ph.D. program ranking 0005∗∗∗ 00009

400015 4000055Father’s education—4-year college or higher 0005∗∗∗ 0002

400025 400015Mother’s education—4-year college or higher 0003 0001

400025 400015Number of observations 3,707 3,002Pseudo R-squared 0012 0004Log-likelihood −21218050 −804090

Source. Authors’ estimation using SDR samples of restricted SESTAT datafor 1995, 1997, and 1999.Notes. The dependent variable is a dummy variable that takes 1 if a scientistis a basic scientist and 0 if he or she is an applied scientist. Years to completefirst bachelor degree is defined as negative of years spent in completing firstbachelor degree (e.g., 4-year is coded as −4). A marginal increase in thisvariable means less time of completion of 1st bachelor degree. A marginalincrease in the variable of Ph.D. program ranking means a higher ranking of aPh.D. program. Estimated marginal effects evaluated at sample mean valuesare reported in the table. Numbers in the parentheses are standard errors.Dummies of demographic characteristics and a field of study are includedin the regression, though these estimates are not reported.

∗∗∗, ∗∗ and ∗ indicate that coefficients are significant at the 1%, 5%, and10% levels, respectively.

regression techniques. Specifically, we first use pooledcross-sectional data to estimate

ln4ygi5 = �0g +�1gexperiencegi +�2g4experiencegi52

+�3gXgi + �gi1 (21)

where ygi is self-reported annualized salary for indi-vidual i in career choice g. Here, experience relatesto labor market experience, and X is a vector of con-trol variables described above. In Tables 5(a) and 5(b),we report the analysis using the longitudinal data,clustering standard errors for individuals who appearmultiple times to account for nonindependence ofobservations.7 Further, given the potential of con-founding of labor experience and cohort or periodeffects (e.g., Glenn 2005), we need to ensure thatthe results are not being driven by differences invintages of cohorts (Stephan 1996). Although labormarket experience and cohort effects cannot be com-pletely separated out empirically with nonexperimen-tal data, we can nonetheless examine the extent towhich cohort effects affect our main findings. InTables 6 and 7, we use panel data for the subsam-ple of repeated observations of scientists. We defineseven cohorts of scientists and engineers based ondoctoral degree completion during a given five-yearperiod, starting from the 1970–1974 cohort and end-ing with the 2000–2006 cohort. Further, we createlabor experience categories of five-year spans (start-ing with less than 5 years and up to greater than25 years). We then perform ordinary least squares(OLS) regressions for each labor experience category,with the dependent variable as the log of real annual-ized salary and independent variables as cohort dum-mies and demographic control variables used above.We also estimate a slope of an earnings equation—theeffect of labor experience—separately for each cohort,and report the analysis for the middle cohorts ofthe 1985–1989 and 1990–1994 graduates. Labor mar-ket experience in the former cohort varies 5–24 yearsand in the latter cohort from 0 to 19 years. Althoughthey are not reported because of space constraints, weobtained similar results for other cohorts.

4. Results4.1. Propositions Related to Career

Choices of ScientistsThe estimated marginal effects for Proposition 1 arereported in Table 3. We test whether proxy variablesfor research ability result in sorting of basic or appliedresearch among academic and industrial scientists,

7 The results are robust to specifications where we use only oneobservation per individual.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 4 Test for Proposition 2 (Choice of Academia vs. Industry Based on Pecuniary and Nonpecuniary Returns)

Panel (a) Panel (b) Panel (c)

(I) (II) (I) (II) (I) (II)

Salary −0012∗∗∗ −0009∗∗∗

400015 400015Benefits 0001 0001

400015 400015Job security 0008∗∗∗ 0006∗∗∗

400015 400015Job location −00005 00001

400015 400015Opportunity for advancement 0004∗∗∗ 0004∗∗∗

400015 400015Intellectual challenge 0001 0004∗∗

400025 400025Level of responsibility −0006∗∗∗ −0004∗∗∗

400015 400015Degree of independence 0010∗∗∗ 0012∗∗∗

400015 400015Contribution to society 0011∗∗∗ 0011∗∗∗

400015 400015Nonpecuniary factors 0013∗∗∗ 0020∗∗∗ 0007∗∗∗ 0009∗∗∗

400025 400025 4000065 4000065Pecuniary factors −0006∗∗∗ −0004∗∗∗ −0001∗ −00003

400015 400015 4000065 4000065Real annual salary ($1,000) −00006∗∗∗ −00007∗∗∗ −00007∗∗∗

4<000015 4<000015 4<0000015No. of observations 7,466 7,466 7,466 7,466 7,466 7,466Pseudo R-squared 0006 0016 0003 0014 0003 0015Log-likelihood −41769008 −41264031 −41932072 −41381060 −41903030 −41341052Inclusion of real salary No Yes No Yes No Yes

Source. Authors’ estimation using restricted SDR data for 2001 and 2003.Notes. The dependent variable is a dummy variable that takes 1 if a scientist is an academic scientist and 0 if he or she is an industrial scientist. In panel (a),each variable of job importance is four-point scale, coded as 4 if very important, 3 if somewhat important, 2 if unimportant, and 1 if not important at all.In panel (b), the variable of nonpecuniary factors is constructed from taking the average of the original variables of challenge, responsibility, and independence.Similarly, the variable of pecuniary factors is constructed from taking average of the original variables of salary and benefit. In panel (c), the variables ofnonpecuniary and pecuniary factors are predicted values from the factor analysis by Bartlett scoring method. The nonpecuniary factor is highly correlated withchallenge, responsibility, and independence. Similarly, the pecuniary factor is highly correlated with salary and benefit. Estimated marginal effects evaluatedare reported in the table. Numbers in the parentheses are standard errors.

∗∗∗, ∗∗ and ∗ indicate that coefficients are significant at the 1%, 5%, and 10% levels, respectively.

respectively. Most demographic variables are not sta-tistically significant. The estimated results show sup-port for Proposition 1: for academic scientists, theestimated marginal effects of grant reception, Ph.D.program ranking, and father’s education are posi-tive and statistically significant at the 1% significancelevel, although this is not the case for industrial sci-entists. For academia, the probability of being basicover applied is likely to increase when scientists spentless time to complete the first baccalaureate degree,received a grant during doctoral study, and had higherparental education levels. By contrast, in industry, esti-mated effects of a grant reception and parents’ educa-tion are statistically insignificant, and time to completea bachelor’s degree has an estimated negative impacton the career option of a basic scientist.

To test Proposition 2 regarding importance ofnonpecuniary benefits, we use survey respondents’

evaluation of the importance of nine attributes oftheir job. Because no universal way to classify a jobattribute to either pecuniary or nonpecuniary benefitsexists, we use three related measures to test Proposi-tion 2. Table 4, panel (a) reports the marginal effectson the likelihood of having an academic career (basicor applied). First, we use all nine original variablesof job importance independently. According to model(I) in panel (a) the probability of being an academicrather than an industrial scientist decreases whensalary or responsibility is rated as an important jobaspect. In contrast, independence, job security, andcontribution to society have a positive effect (similarin magnitude to salary) on the likelihood of choos-ing academia. To address the concern that receivinga high salary correlates with the likelihood of rat-ing it as an important attribute, we add real annualsalary in model (II) of Table 4, panel (a) (and also

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 5(a) Tests for Propositions 3, 4, and 5 (Earnings Profile forEach Career Option) OLS

Academia Industry


Labor experience 0006∗∗∗ 0004∗∗∗ 0004∗∗∗ 0003∗∗∗

4000025 4000035 4000075 4000035Labor experience −00001∗∗∗ −00001∗∗∗ −00001∗∗∗ −00001∗∗∗

squared 4<000015 4<000015 4<000015 4<000015Gender dummy 0017∗∗∗ 0014∗∗∗ 0011∗∗∗ 0005∗∗∗

400015 400025 400055 400025Marriage dummy 0008∗∗∗ 0009∗∗∗ 0002 0005∗∗∗

400015 400025 400045 400025White dummy 0001 0005∗∗∗ −0003 0007∗∗∗

400015 400015 400035 400015Citizenship dummy 0018∗∗∗ 0023∗∗∗ 0004 0005∗∗∗

400025 400025 400045 400025Constant 9077∗∗∗ 9097∗∗∗ 10061∗∗∗ 10071∗∗∗

400025 400035 400065 400035Number of 12,370 8,084 1,194 12,128

observationsR-squared 0031 0016 0010 0005

Source. Authors’ estimation using restricted SDR data for 1995–2006.Notes. The dependent variable is logarithm of annualized earnings. All regres-sions use pooled data of restricted SDR data for 1995, 1997, 1999, 2001,2003, and 2006. Numbers in the parentheses are robust standard errorsclustered at individual level.


Table 5(b) Tests for Propositions 3, 4, and 5 (Equality of EstimatedCoefficients in Table 5(a))

Within institution Between institution

Basic vs. Applied Academia vs. Industry

Academia Industry Basic Applied

(i) Constant <00001 0.15 <00001 <00001(ii) Labor experience <00001 0.15 0002 00001(iii) Labor experience 0035 0.35 0090 0024

squared(ii) and (iii) <00001 0.14 <00001 <00001(i), (ii), and (iii) <00001 0.17 <00001 <00001

Source. Authors’ estimation using restricted SDR data for 1995, 1997, 1999,2001, 2003, and 2006.Note. Numbers in the tables are p-values for testing the hypothesis that thecoefficients are equal across two comparison groups.

in Table 4, panels (b) and (c)). The results remainlargely unchanged and are strengthened by the find-ing that actual salary received results in a lower like-lihood of choosing an academic career. Second, weuse factor analysis of the nine attributes to clusterand account for underlying factors that may systemat-ically influence respondents’ answers to survey ques-tions about job attributes. Using the customary 0.7 offactor loading as a cutoff, the factor analysis suggeststhat salary and benefits may be bundled; accordingly,we use the average value of these two variables to

construct a new variable that captures the importanceof pecuniary benefits. Similarly, the factor analysissuggests that intellectual challenge, responsibility, andindependence can be bundled together, and we inter-pret the average value of these variables as indicatingthe importance of nonpecuniary benefits. Regressionresults using these new variables are reported inTable 4, panel (b). Finally, we predict values of thetwo factors by using the Bartlett scoring method andreport results from these variables in section of thetable. All models show that the more important non-pecuniary benefits are relative to pecuniary returns,the more likely scientists are to sort into academiathan into industry. We note also that our classificationof variables accords with extant literature. For exam-ple, Heneman and Schwab (1985) classify “benefits”as a pecuniary return. Our literature review also sug-gests that “opportunity for advancement” is a grayarea: on one hand, promotions come with monetaryreturns (Lazear and Rosen 1981); on the other hand,advancement allows for nonpecuniary returns such asgreater control/authority (Rynes et al. 1983). In gen-eral, “job security” is nonpecuniary (e.g., Hertzberget al. 1959). People are willing to sacrifice some mon-etary returns to secure a job. Taking the literaturereview and regression results into account, we con-clude that scientists who value nonpecuniary benefitshighly are drawn to academic research. Thus, Propo-sition 2 is supported by the data.

4.2. Propositions Related to Evolution of EarningsTests for Propositions 3–5 using longitudinal data arereported in Tables 5(a) and 5(b). Table 5(a) providesthe OLS estimation results for earnings based on labormarket experience for each career option. Table 5(b)provides the statistical tests for differences in coeffi-cient values across career options for the estimatedearnings profile, with the null hypothesis that eachor a combination of estimated coefficients (constant,labor experience, and labor experience squared) is thesame between any relevant two options. In Figure 2,we provide the estimated earnings trajectories, basedon the coefficients of the linear and quadratic termsof labor experience and at sample mean values of allother independent variables.

The results of the control variables are consistentwith past labor economics studies: regardless of thechosen career option, married, white, U.S. citizen, andmale scientists on average earn more than their coun-terparts, and the estimated earnings profiles are con-cave with respect to labor market experience (Stephan1996). The earnings trajectories are consistent withProposition 3; the average earnings profile of bothbasic and applied academic scientists are lower thanfor industrial scientists (see Figure 2 and tests ofequality in Table 5(b)). Consistent with Proposition 4,

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Figure 2 Estimated Earnings Profiles for Each Career Option

10.0

10.2

10.4

10.6

10.8

11.0

11.2

11.4

11.6

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Log

earn

ings

Years of labor experience

Basic science in academia

Applied science in academia

Basic science in industry

Applied science in industry

Source. Authors’ estimation using restricted SDR data for 1995, 1997, 1999, 2001, 2003, and 2006.Note. Dotted lines represent 95% confidence intervals.

the initial earnings of academic basic scientists arelower than the initial earnings of academic appliedscientists. Indeed, academic basic scientists make thelowest initial earnings of all other scientists, as is evi-dent from Figure 2, tests reported in Table 5(b), andalternative analysis in the online appendix, limitingthe sample to observations for labor experience ofless than or equal to two years. However, consistentwith Proposition 5(a), academic basic scientists alsohave the steepest slope, such that the earnings of aca-demic basic scientists “catch up” to the earnings ofthe industrial scientists toward the very end of thecareer. The coefficient values for academic basic scien-tists are significantly different from the coefficient val-ues for academic applied scientists. Consistent withProposition 5(b), there is no such statistically signifi-cant relationship among industrial scientists for basicand applied research; the null hypotheses for equalityof coefficients cannot be rejected for any combinationof the constant, linear, and quadratic terms of laborexperience.

We also test the estimations of differences in careeroptions using longitudinal data, particularly to inves-tigate whether the results are driven by cohort effects.Given space constraints, we limit our analysis ofProposition 5 to academia, because our theory pre-dicts differences in basic and academic science only

for academic scientists, and not for industrial scien-tists. Table 6 reports the OLS regressions for basic andapplied research, respectively. The dependent variableis the log of real annualized salary within each laborexperience category, and the independent variablesinclude cohort dummies and demographic controlvariables used for cross-sectional regression analyses.The cohort effects in Table 6 are statistically insignifi-cant at conventional significance levels in most experi-ence categories, regardless of basic or applied research.We also estimate the relationship between earningsand labor experience separately for the 1985–1989 and1990–1994 cohorts. Consistent with Proposition 5, inboth cohorts, the estimated slope of an earnings equa-tion in Table 7 is steeper for academic basic scientiststhan for academic applied scientists for the relevantrange of labor experience. Further, in both cohorts,consistent with Propositions 3 and 4, the estimatedcoefficient of the intercept term for industrial scientistsis higher than for academic scientists; within academia,it is higher for applied than for basic scientists (statis-tical tests available on request). Thus, taken together,cohort effects reveal minimal impact on the observedpattern of earnings evolution (Table 6); further, track-ing specific cohorts (Table 7) yields a similar pattern tothe longitudinal analysis.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 6 Tests for Cohort Effects in Academic Basic and Applied Research

Academic basic research Academic applied research

Labor experience categories Labor experience categories

< 5 years 5–9 years 10–14 years 15–19 years 20–24 years < 5 years 5–9 years 10–14 years 15–19 years 20–24 years

Cohort 1 Base Base(1970–1974 graduates)

Cohort 2 Base −0004 Base −0006(1975–1979 graduates) 400065 400075

Cohort 3 Base 00008 00013 Base −0001 −0020(1980–1984 graduates) 400045 400075 400045 400205

Cohort 4 Base −0003 −00009 −0016 Base −0001 0008 0005(1985–1989 graduates) 400045 400075 400175 400045 400055 400205

Cohort 5 Base −0008∗∗∗ −0005 −0003 Base −0016∗∗∗ −0002 −0015(1990–1994 graduates) 400035 400055 400115 400045 400055 400115

Cohort 6 −00007 −0005 −0026∗∗ −0003 −0004 −0025(1995–1999 graduates) 400035 400065 400115 400045 400055 400175

Cohort 7 0005 −0027∗∗ −0007 −0025(2000–2006 graduates) 400035 400115 400085 400165

Constant 10022∗∗∗ 10049∗∗∗ 10078∗∗∗ 10060∗∗∗ 10044∗∗∗ 10024∗∗∗ 10075∗∗∗ 10074∗∗∗ 11002∗∗∗ 11026∗∗∗

400045 400055 400095 400175 400675 400065 400105 400105 400095 400215No. of observations 1,567 1,135 910 793 600 693 581 449 418 322p-value <00001 <00001 <00001 00001 0009 00005 <00001 <00001 <00001 0033R-squared 0002 0006 0005 0003 0006 0002 0004 0005 0005 0003

Source. Authors’ estimation using restricted use SDR data for 1995–2006.Notes. The dependent variable is logarithm of annualized earnings. Numbers in the parentheses are standard errors. The p-values are for a joint significancetest.

∗∗∗, ∗∗ and ∗ indicate that coefficients are significant at the 1%, 5%, and 10% levels, respectively.

5. Model Extensions, AlternativeExplanations, Robustness Checks,and Limitations

5.1. Implications of Relaxation of Assumptionsand Limitations of the Theoretical Model

Several simplifying assumptions enabled us to focusattention on the salient features of sorting of sci-entists and the resultant earnings evolution. Chiefamong them are the assumptions of no uncertaintyand no change across careers to make the model ana-lytically tractable, which prevented us from studyingscientists’ sequential choices of careers. The transi-tion matrices in the SDR database reveal that 97%do not switch careers between industry and aca-demic settings, and 86% do not switch between basicand applied science (Agarwal and Sonka 2010). Fur-ther, there is no significant difference in the fractionof junior scientists in the sample (within two yearsof graduation) doing applied versus basic researchwithin each institutional setting: 16.8% versus 18.5%in academia and 11.1% versus 14.9% in industry.Nonetheless, this simplifying assumption precludesthe possibility that the complementarity betweenbasic and applied scientists in academia increases asthey become more senior and that scientists makecareer choices strategically to maximize their lifetimeutility (Dasgupta and David 1994). Relaxation of this

assumption leads to a revised model that predicts a“cash out” story: basic academic scientists, particu-larly those with higher ability, will switch to appliedscience at later life cycle stages. In addition, the modelsuggests that only able scientists can actually pursuestrategic career switches, because such options are oflimited availability and an assortative matching likelyprevails. This is consistent with the literature on starscientists being more likely to engage in applicationsof their scientific discoveries (e.g., Stuart and Ding2006). Because our model does not allow scientists tomake such career changes, this results in underevalu-ating the earnings evolution of able scientists.

Further, the assumptions that scientists work ineither basic or applied science but not both and thattaste for science is exogenous abstract away fromwithin-individual complementarity of the two activi-ties (Mansfield 1995) and the potential that taste forscience may be endogenous to time allocation of ascientist toward basic and applied scientific activi-ties (Levin and Stephan 1991, Thursby et al. 2007).Thus, our approach is unable to address importantissues such as effects of within-individual comple-mentarity in basic and applied science and endoge-nous taste for science on optimal time allocation,evolution of human capital along basic and applieddimensions, and reputation-building strategies andoutcomes. Incorporation of these elements is beyond

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Table 7 Earnings Equation by Cohort

Academia Industry


Cohort of 1985–1989 graduatesLabor experience 0005∗∗∗ 0002∗∗ 0002 0004∗∗∗

400015 4000095 400025 4000095Labor experience −00001∗∗∗ −00001∗∗ −00001∗∗∗ −00001∗∗∗

squared 40000035 40000025 40000065 40000035Constant 10016∗∗∗ 10050∗∗∗ 11014∗∗∗ 10087∗∗∗

400095 400095 400275 400075Number of 1,866 1,243 210 2 2,259

observationsp-value for <00001 <00001 00664 <00001

joint significanceCohort of 1990–1994 graduates

Labor experience 0006∗∗∗ 0005∗∗∗ 0006∗∗∗ 0003∗∗∗

4000055 4000075 400025 4000065Labor experience −00001∗∗∗ −00001∗∗ −00001∗∗ −00001∗∗∗

squared 40000025 40000025 40000075 40000025Constant 9088∗∗∗ 10000∗∗∗ 10055∗∗∗ 10081∗∗∗

400045 400075 400135 400035Number of 2,765 1,829 279 2,733

observationsp-value for <00001 <00001 <00001 00004

joint significance

Source. Authors’ estimation using restricted use SDR data for 1995–2006.Notes. The dependent variable is logarithm of annualized earnings. Theregression specification is the same as the one in Table 5(a), though we onlyreport constant, labor experience, and labor experience, squared. Numbersin the parentheses are standard errors. For a joint significance test, the nullhypothesis is that all coefficients are zero.


the scope of this paper, given additional complexityand model intractability. Thus, our analysis focuses onwhere scientists conduct research (similar to Aghionet al. 2008 and Stern 2004, who also treat taste forscience as exogenous), assuming that there are noswitches or gradual shifts over time between basicand applied research.

Another simplifying assumption relates to the abil-ity parameters of basic and applied scientists enter-ing symmetrically into the knowledge productionfunction for type B firms in industry. Relaxing thisassumption increases the complexity of the model sig-nificantly and obscures the model’s prediction regard-ing similarity of earnings for industry basic andapplied scientists. The prediction hinges on the factthat the asymmetric force tends to create a gapin earnings profiles of basic and applied scientists,whereas the synergy and complementary forces workto narrow that gap. However, for a reasonable rangeof parameters, we can still obtain the model pre-diction. Further, the assumption of stable matching(and related assumption of full information) helpedderive an equilibrium matching pattern. As in Shimer(2005), when a coordination friction is introduced at

the expense of analytical complexity, the assortativematching becomes imperfect. Some less able scien-tists are paired with more capable research institu-tions or scientists, but the average characteristic of asorting pattern and earnings evolution remains thesame qualitatively.

In the context of the earnings profile, relaxing theassumption that scientists’ pay is not linked to cur-rent investment in human capital but only to accu-mulated human capital permits the possibility thatcurrent human capital investments also create imme-diate returns and will result in more time devotedto human capital investment, provided that scien-tists’ other activities are assumed away. This makesthe earnings profiles steeper than currently predicted.Alternatively, if one models the scientists’ other activ-ities as requiring both time and current researchoutput, human capital accumulation slows down,resulting in a flatter earnings profile, though becauseof shift up because of income from other activities.

5.2. Robustness Checks, Alternative Explanations,and Limitations for the Empirical Findings

Our empirical analysis used a large sample of scien-tists and therefore provided general empirical insightsregarding the model’s predictions. However, severalalternative explanations need to be addressed, par-ticularly as they relate to underlying heterogeneityof jobs and characteristics of scientists, resulting insimilar predictions. We first examine whether the dif-ferences in earnings profile between academic andindustry scientists may be a result of gender com-position: robustness checks reported in the onlineappendix reveal that the propositions are supportedfor both male and female subsamples. A second con-cern is whether mean estimates of earnings pro-files from OLS regressions are consistent with thedata at different points of the earnings distribution—i.e., it is possible that opposite effects of high earn-ers and low earners cancel each other out to generatethe average earnings profiles estimated from OLSregressions. Quantile regressions at the first, second,and third quartiles show that the model’s predic-tions regarding earnings profiles are supported (seethe online appendix). Furthermore, this result indi-cates that the assortative sorting story is consistentwith the data and strengthens our empirical analy-sis. Third, we examine whether the propositions arerobust to within-science field comparisons for lifeand physical sciences, particularly because the delin-eation of “basic” versus “applied” may differ acrossthem. We could not conduct this analysis for otherscience and engineering areas, because of the lackof sufficient observations for meeting NSF disclosurerequirements. As seen in the online appendix, withone exception, all the propositions are largely sup-ported within the subsamples. The only proposition

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


for which we lacked support was Proposition 4 in thephysical sciences; we did not find support for differ-ences in initial earnings between academic basic andapplied scientists. We also note that in the life sci-ences, we received partial support for Proposition 5:in addition to the predicted steepness of the earningsslope for basic scientists relative to applied scientistsin academia, we also find similar results in industry(as opposed to the predicted no significant differencein slopes in industry). Consistent with extant studiesthat use biotechnology as an exemplar of an indus-try that is based on basic science, the results indicatethat basic scientists, regardless of whether they are inacademia or industry, enjoy higher returns relative toapplied scientists. In unreported regressions, we alsoconfirmed that our analysis is robust to the exclusionof post doctoral scientists and of nontenure track fac-ulty in academia.

Among the empirical limitations of our study, chiefis our inability to rule out some competing expla-nations. Nonetheless, we are able to examine theexplanatory power of alternative models relative toours by using our model’s unique implications asguidance. In particular, we believe that although someof the results may be explained piecemeal by alter-native theories, these theories cannot simultaneouslyaccount for all the empirical findings. For example,earnings differentials between academia and industrycould be attributed to a relatively small supply of aca-demic positions, but this simple supply-and-demandstory cannot explain similarities and/or differencesin the evolution of earnings profiles between basicand applied scientists. High fixed costs in appliedscience may induce research institutions to hire ablescientists and lead to higher earnings of applied sci-entists, but they cannot explain differences in theslopes observed across institutional settings, partic-ularly when compared with the earnings slopes ofbasic scientists. Similarly, differences in preferencesfor nonpecuniary benefits between basic science andapplied science within each research institution mayresult in earnings differentials between the two, butwould not explain why the differential exists withinacademia and not industry. Alternatively, as Nicker-son and Zenger (2008) argue, peer envy may result insimilar compensation for basic and applied scientistsin industry, but this explanation leaves unansweredwhy academic settings are free from such peer envy.More importantly, it is hard to explain the evidenceof this study that salary is a more important factor forindustrial scientists than academic scientists.

Although the empirical results of this study arelargely consistent with the theoretical implications,we acknowledge that our empirical analysis isdescriptive in the sense that it explores empirical pat-terns of sorting and earnings evolution rather than

establishing a causal relationship. Accordingly, someempirical findings in this paper need to be interpretedwith caution. Further, we note that our main interestis not to estimate an earnings profile of each careerpath for a scientist randomly chosen from the popula-tion; thus, we do not attempt to control for potentialselection effects in the empirical analysis. Nonethe-less, our theoretical model regarding the sorting ofscientists guides us to infer a direction of the poten-tial selection effects in our empirical results. Becausethis and other studies report that academic scien-tists have a taste for science, an estimated earningsprofile of academic scientists in our study would bedownward biased from random selection. Ability orlearning ability sorting may result in an upward biasfrom random assignment for the estimated slope ofbasic scientists in academia and a downward bias forapplied scientists in academia. An estimated earn-ings profile of industrial scientists is likely upwardlybiased from random selection, given the positive sort-ing between basic and applied scientists. Finally, weare also cognizant of the fact that additional usefulinsights may be obtained by empirically examininga causal or structural relationship between sortingof scientists and earnings evolution. Unfortunately, asimple self-selection model such as a Roy model (Roy1951, Heckman and Honoré 1990) does not resolve aself-selection problem because of the nature of two-sided matching. The problem is further exacerbatedby the fact that we do not have data from a ran-dom experiment. Accordingly, we leave it for futureresearch to firmly establish a causal relationship bydeveloping a full structural estimation model with atwo-sided matching estimation.

6. Discussion and ConclusionOur paper integrates matching theory with life cyclemodels of human capital investments to analyze labormarkets for scientists, where heterogeneous character-istics of scientists and institutions result in an optimalmatch between specific individuals and the alterna-tive career options. Specifically, we model how scien-tists who differ across two dimensions—ability andtaste for nonpecuniary returns—choose their careers.The sorting patterns for early career scientists, alongwith differences in complementarities between basicand applied scientists and in investments in humanand physical capital within academia and industryalso have implications for the earnings trajectory forscientists within each career option. We test the modelimplications using comprehensive and rich data com-piled by the NSF’s SESTAT program and find supportfor its predictions.

Our model’s predictions for the sorting of scien-tists and the resultant earnings evolution contributeto the literature in the economics of science. By inte-grating insights from matching theory (Becker 1973)

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


into traditional models of scientific labor marketsthat build on life cycles in human capital invest-ments (Becker 1962), we address Stephan’s (1996) callfor better modeling of the economics of science. Forexample, some of our model implications are veryconsistent with the implications generated by lifecycle models (e.g., Levin and Stephan 1991, Thursbyet al. 2007), but incorporating the two-sided match-ing element permits us to generate new implications:our study models how competition within scientistsfor preferred jobs and within institutions for preferredscientists results in matching based on ability, pref-erences, and complementarity in multifactor scientificproduction functions. The matching outcomes in turninfluence the evolution of human capital and earn-ings over time. Our model thus generates some novelinsights that are backed by the empirical analysis: inacademia, scientists of higher ability sort into basicrather than applied research, and initial earnings ofbasic scientists are lower but the slope of earningsis higher relative to applied scientists. In industry,by contrast, there is no such ability sorting, and theearnings trajectories of basic and applied scientistsare similar to each other. Thus, we hope that ourstudy sheds light on how possible self- or sample-selection biases may influence applications of the lifecycle model to specific empirical contexts.

Our study also contributes new insights by build-ing on extant work related to the “taste for sci-ence” (Roach and Sauermann 2010, Stern 2004). Wepredict and find that a higher taste for nonpecu-niary returns sorts scientists in academia over indus-try. Stern (2004) focuses primarily on industry andshows that scientists with higher “taste for science”are willing to accept a compensating wage differen-tial to work for science oriented firms, and we usea similar logic to show how such preferences mayresult in these scientists preferring academia overindustry. Further, by incorporating life cycle–relatedtrade-offs between current and future opportunities,we generate additional important insights regard-ing the relation between a taste for nonpecuniaryreturns and earnings trajectories. Our study revealsthat the initial earnings of basic scientists are lowerthan applied scientists within academia, in spite ofthe fact that science norms and nonmonetary benefitsare very similar in these research domains. We positand show that basic academic scientists may sacri-fice current earnings for steeper growth over theirlife cycle, relative to applied academic scientists. Theempirical analyses support our model prediction thatbasic and applied scientists have similar earningsin industry, given synergies between the two. Thus,our model implications indicate that the observedearnings differential between academia and industrystem from both differences in a taste for science and

the presence (absence) of complementarities betweenbasic and applied scientists in industry (academia).This result holds even after controlling for observableand unobservable characteristics of scientists.

Further, extant work on careers of scientists typi-cally examines either industry or academic scientistsor examines the institutional differences in nar-rowly defined fields such as life sciences (Bercovitzand Feldman 2008, Roach and Sauermann 2010,Sauermann and Stephan 2012, Stuart and Ding 2006).Using a broad, rich data set allows us to build onextant work by comparing industry and academic sci-entists across a broad, generalizable range of fields.Our model and empirical findings are consistent withthe survey-based study of Roach and Sauermann(2010), who show that scientists self-select betweenacademia and industry based on their “taste for sci-ence,” and with Sauermann and Stephan (2012), whoexplore similarities and differences between academicand industrial science. Consistent with Roach andSauermann (2010) and our study’s Propositions 2and 3, Sauermann and Stephan (2012) find that aca-demic scientists are relatively more satisfied withnonpecuniary benefits than industrial scientists, thusunderscoring the importance of nonmonetary benefitsin sorting scientists between academia and industry.Further, our longitudinal findings related to Proposi-tions 4 and 5 regarding earnings profiles within andacross institutional settings provide a more nuancedrelationship that complements their cross-sectionalfindings. They find that academic scientists earn lessthan industrial scientists; we find a striking similarityof earnings profiles between basic and applied scien-tists within industry, but divergence of earnings pro-files between basic and applied scientists in academia.Additionally, we find that at later career stages, basicacademic scientists’ earnings are not significantly dif-ferent from those of industrial scientists.

Our study provides managerial and policy impli-cations for each institutional setting that we exam-ine. Within industry, policy makers and innovationmanagers need to recognize that high complementar-ities between basic and applied scientists permit themto attract equally able scientists in either researchdomain, and thus these complementarities deserve tobe emphasized by creating cultures that recognize theimportance of basic science to technological applica-tions. Within academia, our study offers some insightsto policy makers and administrators, particularly inthe hiring and retention policies for scientists in basicand applied domains. For instance, our model showsthat a sorting of ability within basic and applied aca-demic research is driven by the differential invest-ments in complementary physical capital. Thus, tothe extent that national science and innovation policy

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


may encourage universities to engage in more tech-nology transfer, there is a need to develop strongercomplementarities between basic and applied scien-tists. Existing models underscore the value of timeallocation of an individual scientist between basicand applied activities, but our study highlights thatincreasing complementarities across individuals thatspecialize in basic or applied science may also havebeneficial outcomes.

Our study also offers young scientists seeking toembark on alternative careers an understanding of thesorting patterns and average earnings, so that theycan make informed decisions based on an assessmentof their own ability and preferences. This is even moreimportant because all young scientists are trained inacademia, where there is often a pejorative assess-ment placed on either applied research or working inindustry settings (Stephan 1996). Given that only 26%of scientists are employed in basic academic research(Agarwal and Sonka 2010), for the United States toremain competitive in an increasingly knowledge-based economy, there is also a need for renewedassessment of our science education policy, and toexamine if our universities provide the other 74% oftheir graduate student population with the necessaryknowledge, skills, and attitudes required for success.Such an assessment may lead to the more widespreaddevelopment of innovative programs within universi-ties that will provide young scientists with the literacyand experiential learning of the economic, business,and legal issues they may encounter in their careers(Agarwal and Sonka 2010). Similarly, national inno-vation policies may also want to encourage focusedprograms developed for lifelong learning skills forscientists and provide for the development of con-tinued education programs that permit scientists inthe workforce to efficiently learn the complementarymanagement and entrepreneurship skills that theyneed to be more successful in leveraging their scien-tific accomplishments.

In summary, our theoretical model and the empir-ical findings identify the impact on institutional dif-ferences on the microlevel choices made by economicagents (firms and universities on the demand side,and scientists with heterogeneous ability and pref-erences on the supply side) and is an importantfirst step toward identifying the extent to whichmarket forces guide the allocation of scientists. Ourtheoretical study demonstrates this point by usinginsights from two-sided matching theory. Based onthe premise that there may be higher comple-mentarities between basic and applied scientists inindustry production functions relative to academia,our model derives distinctive implications regardingcareer choices and resultant earning trajectories of sci-entists. As a result, our study can be regarded as

a minimal test (i.e., a necessary condition test) forthe existence of the complementarity and contributesto the literature that has previously documented thesynergies in basic and applied research conducted inindustry settings.

AcknowledgmentsBoth authors contributed equally to this paper, and theirnames are in alphabetical order. The authors appreciatecomments received from the editors and two anonymousreviewers, Serguey Braguinsky, Benjamin Campbell, SethCarnahan, Raj Echambadi, Martin Ganco, Denisa Mindruta,Mahka Moeen, Henry Sauermann, Paula Stephan, PeterThompson, Marie Thursby, and seminar participants atthe 2009 REER and 2010 Maryland Entrepreneurship con-ferences. All remaining errors are the authors’ responsi-bility. The authors gratefully acknowledge grant supportfrom the Ewing Marion Kauffman Foundation and theNational Science Foundation (NSF). The empirical analysisuses restricted access data from the NSF, which does notimply NSF endorsement of the research methods or conclu-sions contained in this paper.

ReferencesAgarwal R, Sonka S (2010) Different strokes for different folks:

University programs that enable diverse career choices ofyoung scientists. Libecap G, Thursby GM, Hoskinson S, eds.Advances in the Study of Entrepreneurship, Innovation and Eco-nomic Growth, Spanning Boundaries and Disciplines: Uni-versity Technology Commercialization in the Idea Age, Vol.21 (Emerald Publishing, Cambridge, MA), 139–164.

Aghion P, Dewatripont M, Stein JC (2008) Academic freedom,private-sector focus, and the process of innovation. RAND J.Econom. 39(3):617–635.

Argyres N, Liebskind J (1998) Privatizing the intellectual com-mons: Universities and the commercialization of biotechnologyresearch. J. Econom. Behav. Organ. 35(4):427–454.

Becker GS (1962) Investment in human capital: A theoreticalapproach. J. Political Econom. 79(5):9–49.

Becker GS (1973) A theory of marriage: Part I. J. Political Econom.81(4):813–846.

Ben-Porath Y (1967) The production of human capital and the lifecycle of earnings. J. Political Econom. 75(4):352–365.

Bercovitz J, Feldman M (2008) Academic entrepreneurs: Organiza-tional change at the individual level. Organ. Sci. 19(1):69–89.

Boardman PC, Bozeman B (2007) Role strain in university researchcenters. J. Higher Ed. 78(4):430–463.

Boardman PC, Ponomariov BL (2007) Reward systems and NSFuniversity research centers: The impact of tenure on univer-sity scientists’ valuation of applied and commercially relevantresearch. J. Higher Ed. 78(1):51–70.

Braxton JM, Luckey W, Helland P (2002) Institutionalizing a BroaderView of Scholarship Through Boyer’s Four Domains (ASHE-ERICHigher Education Report). (Association for the Study of HigherEducation, Washington, DC).

Bush V (1945) Science the Endless Frontier: A Report to the Presidenton a Program for Postwar Scientific Research (U.S. GovernmentPrinting Office, Washington, DC).

Cohen WM, Levinthal DA (1990) Absorptive capacity: A newperspective on learning and innovation. Admin. Quart. Sci.35(1):128–152.

Dasgupta P, David P (1994) Toward a new economics of scienceand technology. Res. Policy 23(5):487–521.

Etzkowitz H (1998) The norms of entrepreneurial science: Cogni-tive effects of the new university-industry linkages. Res. Policy27(8):823–834.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.


Geisler E (1989) University-industry relations: A review of majorissues. Link AN, Tassey G, eds. Cooperative Research andDevelopment: The Industry-University-Government Relationship(Kluwer, Boston), 43–64.

Gittelman M, Kogut B (2003) Does good science lead to valuableknowledge? Biotechnology firms and the evolutionary logic ofcitation patterns. Management Sci. 49(4):366–382.

Glenn N (2005) Cohort Analysis (Sage Publications, Thousand Oaks,CA).

Golderberger ML, Maher BA, Flattau PE (1995) Research DoctoratePrograms in the United States; Continuity and Change (NationalResearch Council, Washington, DC).

Goldfarb B (2008) The effect of government contracting on aca-demic research: Does the source of funding affect scientific out-put? Res. Policy 37(1):41–58.

Griliches Z (1986) Productivity, R&D, and basic research at the firmlevel in the 1970s. Amer. Econom. Rev. 76(1):141–154.

Haveman R, Wolfe B (1995) The determinants of children’s attain-ments: A review of methods and findings. J. Econom. Literature33(4):1829–1878.

Heckman JJ, Honoré BE (1990) The empirical content of the Roymodel. Econometrica 58(5):1121–1149.

Heneman HG III, Schwab DP (1985) Pay satisfaction: Its mul-tidimensional nature and measurement. Internat. J. Psych.20:129–141.

Hertzberg F, Mausner B, Snyderman BB (1959) The Motivation toWork (John Wiley & Sons, New York).

Hounshell DA, Smith JK (1988) Science and Corporate Strategy(Cambridge University Press, New York).

Kremer M, Maskin E (1996) Wage inequality and segregation byskill. NBER Working Paper w5718, National Bureau of Eco-nomic Research, Cambridge, MA.

Lazear EP, Rosen S (1981) Rank-order tournaments as optimumlabor contracts. J. Political Econom. 89(5):841–864.

Levin SG, Stephan PE (1991) Research productivity over the lifecycle: Evidence for academic scientists. Amer. Econom. Rev.81(1):114–132.

Leibowitz A (1974) Home investments in children. J. PoliticalEconom. 82(2):S111–S131.

Mansfield E (1995) Academic research underlying industrial inno-vations: Courses, characteristics, and financing. Rev. Econom.Statist. 77(1):55–65.

Merton R (1973) The Sociology of Science: Theoretical and EmpiricalInvestigation (University of Chicago Press, Chicago).

Mindruta D (2012) Value creation in university-industry rela-tionships: A matching approach. Strategic Management J.Forthcoming.

Murray F, Stern S (2007) Do formal intellectual property rightshinder the free flow of scientific knowledge? An empiricaltest of the anti-commons hypothesis. J. Econom. Behav. Organ.63(4):648–687.

National Science Board (2010) Science and Engineering Indicators 2010(National Science Foundation, Arlington, VA).

Nickerson JA, Zenger TR (2008) Envy, comparison costs, andthe economic theory of the firm. Strategic Management J.29(13):1429–1449.

Roach M, Sauermann H (2010) A taste for science? Ph.D. scientists’academic orientation and self-selection into research career inindustry. Res. Policy 39(3):422–434.

Rosenberg N (1990) Why do firms do basic research (with theirown money)? Res. Policy 19(2):165–174.

Roth AE, Sotomayor M (1992) Two-Sided Matching: A Study in Game-Theoretic Modeling and Analysis (Cambridge University Press,Cambridge, MA).

Roy A (1951) Some thoughts on the distribution of earnings. OxfordEconom. Papers 3(2):135–146.

Rynes S, Schwab DP, Heneman HG (1983) The role of pay and mar-ket pay variability in job application decisions. Organ. Behav.Human Performances 31(3):353–364.

Sacerdote B (2000) The nature and nurture of economic outcomes.NBER Working Paper 7949, National Bureau of EconomicResearch, Cambridge, MA.

Sattinger M (1993) Assignment models of the distribution of earn-ings. J. Econom. Literature 31:831–880.

Sauermann H, Stephan P (2012) Conflicting logics? A multidimen-sional view of industrial and academic science. Organ. Sci.,ePub ahead of print September 10, http://dx.doi.org/10.1287/orsc.1120.0769.

Shimer S (2005) The assignment of workers to jobs in an econ-omy with coordination frictions. J. Political Econom. 113(5):996–1025.

Stephan PE (1996) The economics of science. J. Econom. Literature34(3):1199–1235.

Stern S (2004) Do scientists pay to be scientists? Management Sci.50(6):835–853.

Stokes DE (1997) Pasteur’s Quadrant (Brookings Institution Press,Washington, DC).

Stuart T, Ding W (2006) When do scientists become entrepreneurs?The social structural antecedents of commercial activity in theacademic life sciences. Amer. J. Sociol. 112(1):97–144.

Thursby JG, Thursby MC (2002) Who is selling the ivory tower?Sources of growth in university licensing. Management Sci.48(1):90–104.

Thursby M, Thursby JG, Gupta-Mukherjee S (2007) Are there realeffects of licensing on academic research? A life cycle view.J. Econom. Behav. Organ. 63(4):577–598.

Copyright:

INFORMS

holdsco

pyrig

htto

this

Articlesin

Adv

ance

version,

which

ismad

eav

ailableto

subs

cribers.

The

filemay

notbe

posted

onan

yothe

rweb

site,includ

ing

the

author’s

site.Pleas

ese

ndan

yqu

estio

nsrega

rding

this

policyto

perm

ission

s@inform

s.org.

industry or academia, basic or applied? career choices and...

Documents