Empirical Analysis of Career Transitionsof Sciences and Engineering Doctorates in the US
Natalia Mishagina∗,Queen’s University †
August 10, 2009
Abstract
This paper studies the career choices and mobility of white male doctorates in sci-ences and engineering (S&E). Relevance of employment to R&D and S&E is evaluatedusing an alternative approach that uses activities on the job rather than occupations.Transitions between tasks were assessed using a duration model with competing risksestimated on the Survey of Doctorate Recipients (1973-2001). It was found R&D tasksemploy only 57% of S&E doctorates, and this number varied over time. Most transi-tions out of R&D happen when doctorates are young. Decisions to switch tasks areaffected by demographics, performance on the job, and general economic conditions.
Keywords: Occupational transitions; duration analysis; competing risks; scienceand technology workforce; high-skilled labor.
JEL Classification Codes: C41, J24, J44.
∗I am grateful to Christopher Ferrall for his guidance in working on this paper. I appreciate commentsand suggestions from Charles Beach, Paula Stephan, Sharon Levin, Stephane Robin, Mark Regets, SusanHill, the participants of the CEA (Montreal, 2006) and SRS/SEWP workshop (Arlington,VA, 2006). Iwould also like to thank SRS, National Science Foundation, for providing the data and valuable assistance.I acknowledge Research Fellowship from the SRS/American Statistical Association.
†Queen’s University, Dunning Hall, Room 209 Kingston, ON, K7L 3N6 Canada Tel: (312) 533-2250.Email: [email protected]
1 Introduction
This paper studies the career choices and mobility patterns of white male PhDs in natural
sciences and engineering (S&E). While a PhD career is typically associated with research and
development (R&D), there is anecdotal evidence that a large fraction of S&E doctorates find
employment opportunities in non-traditional for scientists areas such as financial, accounting,
and other business services. This fact has never been acknowledged or assessed in the relevant
literature before. In fact, little is known about the employment choices of S&E doctorates
and especially their career outcomes once they leave S&E. This raises the following questions:
What do we know about the career choices of S&E doctorates and their relevance to S&E
in general and R&D in particular? How does outward mobility from S&E compare to that
within S&E? And finally, what factors affect the decisions of scientists and engineers to
switch careers and the timing of the change?
The consequences of career transitions depend on their nature and the factors that affect
them. Whether these transitions are of concern for public policy depends on who makes
the transitions, when, and why. For example, if those who leave R&D are the leading
researchers, then one should be concerned with losing highly trained specialists. If, however,
switchers are the relatively “worse” scientists, then it may be worrisome that this sorting did
not happen earlier, say during their training. The reasons for switching also shed light on
how marketable the skills of doctorates are outside their professions. One may suspect that
switchers are lured away from research by better monetary rewards and career advancement
prospects in non-S&E areas. Then the question is: What do non-S&E employers value
in doctorates? Perhaps, employers believe that the ability to generate interesting ideas and
conduct independent research is positively correlated with the ability to perform well in non-
S&E. Perhaps, some aspects of the research skill is transferable to non-S&E; for example,
mathematical and computing skills, modeling and analytical abilities. Alternatively, a PhD
degree is just a signal of exceptional intellectual abilities and suggests a high possibility to
easily re-train a doctorate to meet the demands of a new career.
The first objective of this paper is to assess the relevance of PhD employment to S&E
1
in general, and R&D in particular. The second goal is to compare the employment rates
and mobility patterns of doctorates in different types of careers. The third objective is
to evaluate how various personal and job characteristics, research productivity, and labor
market conditions affect mobility rates and timing within and out of S&E. To do so, an
econometric model of transitions with competing risks is specified and estimated. To meet
these objectives, this paper uses the Survey of Doctorate Recipients (SDR) collected by the
National Science Foundation (NSF) from 1973 to 2001. This longitudinal data set is unique
because of its large representative sample of professionals whose share in the labor force is
small. As a data source on careers, the SDR is rich in employment characteristics specific
to S&E.
To achieve the first goal, I propose a new approach that classifies employment choices
based on primary activities first and only then on occupational titles and employment sectors.
According to this method all individuals in the survey were categorized as employed in one
of three types of employment or “tasks”: 1) creation of new knowledge (“R&D task”), 2) ap-
plication of existing knowledge (“applied tasks”), and 3) tasks unrelated to S&E (“non-S&E
tasks”). This approach is an alternative to a more traditional method that would involve us-
ing occupational titles or sectors of employment alone. According to this traditional method,
only “scientists” or even “academic scientists” would be qualified as employed in R&D. In
this case, a recent decrease in available tenured-track positions in academia would be con-
sidered as a threat to R&D because it reduces employment opportunities. The approach
proposed in this paper allows one to see jobs in R&D from a very different angle. It uncovers
the ever-evolving nature of R&D that has spilled out of academia into other sectors. More-
over, it helps evaluate retention rates in R&D correctly because by construction it is not
sensitive to individuals changing firms or sectors as long as R&D activities remain primary
activities on the job. Finally, it eliminates confusion related to interpretation of occupational
titles that often occur in surveys when occupations are self-reported.
I found that in the data only 57 percent of doctorates worked in tasks directly oriented
towards R&D, including its supervision. Another 35 percent worked in applied tasks, such as
2
software development by non-computer science majors, teaching, or professional services. Fi-
nally, 8 percent of doctorates worked in non-S&E tasks in such sectors as financial consulting
or accounting.
The second finding is that task choices vary within a PhD career. Doctorates tended
to begin their careers in R&D tasks (72%), but only roughly half of them remained there
thirty years later. Some of those who left R&D went to applied tasks (80%), and others
left S&E for good (20%). The transitions within and out of S&E differ in their timing and
patterns. The only common feature was their high frequency within the first 16 years of
task-specific tenure. Decisions to switch tasks did not differ significantly by cohort with
minor exceptions. For example, those who graduated in the 1980s were the most likely to
leave R&D for application in all specifications. At the same time, mobility from R&D to
non-S&E was the lowest for this cohort (1980s). The transitions differed by field of study,
current activity, sector of employment, and academic rank conditional on being in academia.
Analysis of the employment histories showed that mostly academic researchers on tenure-
track or with tenure switched to applied tasks. This finding agrees with the predictions of
the human capital theory that scientists invest into their experience or reputation in order
to cash in on it later on, Cater and Lew (2005). It was also found that individuals who leave
S&E are more likely to be employed in non-academic sectors and have non-S&E degrees
prior to their PhD. Several measures of research productivity and recent research activity
prior to transitions had a significant and positive effect only on the transitions from applied
tasks back to R&D tasks, although low response rates to all productivity questions make
interpretation of the results difficult. Another result is that the decisions to switch tasks are
sensitive to the labor market conditions measured by the overall unemployment rate, share
of the R&D expenditures in the GDP, and enrollment rates in S&E postsecondary programs.
The final interesting result, which is a by-product of the main analysis of the paper,
concerns the nature of postdoctoral appointments and their influence on a doctoral career.
It was found that individuals who held several postdoctoral appointments were more likely
to leave S&E for good, whereas they did not exhibit higher transition rates between tasks
3
within S&E. This supports the hypothesis that postdoctoral appointments are probationary
positions or “waiting lists” of a sort rather than extensions to doctoral education that provide
additional training. This new result contributes to understanding of the market for scientists.
This study contributes to several strands of literature. Mostly, it contributes to the
growing literature on employment choices of high skilled labor and S&E as industry. In
particular, this study proposes a method of distinguishing the jobs of PhDs with respect to
their relevance to R&D from other types of employment. A detailed analysis of employment
choices of S&E PhDs and their dynamics throughout a career is conducted. This study
is also the first to document the employment choices of PhDs outside the scope of S&E.
Results on behavior of postdoctorates described above are also new and deserving further
investigation. Finally, the effect on mobility of such factors as labor market conditions and
research productivity has never been assessed before.
2 Literature
The literature on science and technology as an industry together with its labor force is
well summarized in Stephan (1996). In particular, the author discusses relevant issues for
this paper such as the training of doctoral scientists and engineers, markets for their skills,
incentives, and earnings profiles, the importance of the non-pecuniary benefit of research
(i.e., “ the joy of puzzle solving”), and research funding and its influence on scientists’
productivity.
There is also literature that focuses on the specific aspects of the traditional career paths
for doctorates; for example, R&D or teaching in academia, government or the private sector.
Levin and Stephan (1991) study a very similar problem to the one raised in this paper. The
authors develop and estimate a dynamic model to study how academic scientists allocate time
to research versus teaching. Research in their model has an investment motive: Discoveries
improve the scientists’ stock of knowledge, which in turn enhances scientists’ future research,
and helps in teaching. Teaching is the only activity that provides consumption benefits.
In this study, I analyze mobility between research and teaching activities, which assumes
4
specialization in one of the two rather than the optimal allocation of time between them.
The mobility out of R&D and the complete career changes of doctorates considered in this
paper have not been assessed in the S&E literature before. The studies of traditional career
paths can be grouped into two major strands. The first strand is concerned with employment
and mobility within a specific sector, such as managerial versus technical jobs in R&D firms
(Ferrall (1997), Biddle and Roberts (1994)), promotions within academia (Robin (2002),
Grimes and Register (1997)), or mobility between firms (e.g Moen (2001), Fallick et al.
(2005), Almeida and Kogut (1999)). The second strand studies mobility between R&D
sectors (e.g., Zucker and Darby (2006), and Audretsch and Stephan (1999)).
Mobility within R&D sectors serves as a vehicle of transferring new ideas through their
inventors who most often hold patents on these inventions rather than publish their findings
freely in scientific journals. In these cases employing the inventors becomes the only way
to access new ideas. On one hand, this type of mobility is a concern since it reduces firms’
incentives to invest in human capital. On the other hand, it is beneficial for the recipient
firms. Almeida and Kogut (1999) analyze the mobility of engineers who are patent-holders
and the patterns of their patent citations in the semiconductors industry. The authors find
notable regional differences in knowledge localization, especially in such regions as Silicon
Valley and the New York Triangle. They explain this phenomenon by the existence of a large
local market for scientists and engineers in these regions. Their finding about high mobility
is supported by Fallick et al. (2005) who analyze interfirm transitions in high technology
clusters using the CPS data. They have discovered very high transition rates within computer
clusters in California, such as Silicon Valley. The authors explain this phenomenon with the
features of the state law such as restriction on non-compete agreements. Firms with superior
innovations benefit from such mobility because it causes the reallocation of human resources
towards these firms. The authors show that under certain conditions there is a net benefit
of this kind of mobility, but for technological reasons it is specific to the computer industry
only. Moen (2001) considers this issue using the human capital framework. He shows that in
R&D-intensive firms workers are paid much lower wages early in their career. He considers
5
this a payment for acquiring training, for which they are compensated later in their careers.
He suggests that all possible externalities of labor mobility are internalized by the market.
The second strand of the literature considers mobility between different sectors such as
academia or government laboratories and industry (e.g., Zucker et al. (1997), Zucker and
Darby (2006), Audretsch and Stephan (1999)). This type of mobility is interesting because it
causes knowledge diffusion across sectors through the reallocation of researchers. Zucker et al.
(1997) study behavior of academic bioscientists moving to industry. Later Zucker and Darby
(2006) extend the scope of analysis to all scientific fields. Both studies estimate a duration
model to show that it is the “star” scientists that make the move. The quality of a scientist
is determined by the number of primary accession numbers (similar to citations) obtained
from the GenBank. Audretsch and Stephan (1999) study mobility of doctorate biologists
who left academia to found bioscience firms. They also found that the “switchers” are the
“star” scientists, who start their own company rather than join a bigger one. This interesting
observation proves the prediction of the human capital theory that early in their careers,
academic scientists would heavily invest in their knowledge in order to build reputation, and
cash it out later, by selling their services to a private firm or by starting their own business.
A slightly different aspect of mobility was considered in an earlier paper, by Biddle and
Roberts (1994), where the authors model and empirically test one of the career paths of
industrial engineers that begins in a technical occupation and then can be continued in a
managerial occupation. A self-selection and job matching framework was used to determine
who decides and why to switch to the managerial track and whether to stay on it or not.
The key assumption in the paper is that the skills to perform technical and managerial
tasks are positively but imperfectly correlated. Therefore, the model predicts that the most
productive technical workers would also be more productive as managers, and thus they
choose to switch to management later in their career. Another prediction of the model
was that those who switched tracks and turned out to be unproductive as managers would
backtrack. The authors found empirical support for the first prediction of the model but not
the second.
6
The next strand of the relevant literature is concerned with retention in other professions
and career changes. The following studies are especially closely related to the present paper
because they answer similar questions about the careers of high school teachers. For example,
Stinebrickner (1998) and Stinebrickner (2002) evaluate reasons for teachers attrition from
the profession. Both studies use duration analysis techniques to evaluate the factors that
affect the timing and reasons for exiting the profession. They find that teachers who exit
are more likely to be women who leave the labor force altogether for raising newly-born
children. Scafidi et al. (2007) assesses how attrition is affected by the availability of higher
paid occupations outside teaching. Similarly to doctoral scientists and engineers, teachers’
skills are marketable outside the teaching sector. Non-teaching sectors offer higher salaries
which facilitates the teachers’ mobility out of the profession. However, the authors find that
attrition to higher paid occupations is not common, at least not among the female teachers
they studied.
Two more studies should be mentioned as relevant to the current investigation. The first
is a study by Van den Berg et al. (2002) who compare various paths from medical student to
junior medical specialist (or “trainee”) based on the data from the Netherlands. The authors
use a duration model with competing risks to evaluate the time to become a trainee from
two different origins: unemployment versus medical assistant. They find that it is easier to
become a trainee if an individual chose to be a medical assistant first. Therefore, the job of
a medical assistant may be considered a stepping-stone job. Although the model does not
explain whether employment as a trainee provides additional skill or simply scans for the
most talented or eager to learn students, this finding is interesting and very relevant.
The second study by Sauer (1998) studies early job choices of lawyers. The author de-
velops and estimates a dynamic discrete choice model where employment options differ in
terms of their monetary and non-pecuniary compensations, retention and dismissal proba-
bilities, and their effect on future options. The results show that motives to choose jobs
in the private sector, which are highly demanding and highly paid, differ depending on the
ability of an attorney. High-ability attorneys face high promotion probabilities and high
7
earnings in these jobs so it is natural for them to choose them. Low ability attorneys these
jobs even if their survival probability in these jobs is low. This finding would be puzzling
if the effects of current jobs on future employment was ignored. The author suggests that
employment in private firms early in the career improves the chances of low-ability lawyers
to secure well-paid jobs in other sectors.
3 Empirical evidence on career choices and transitions
This section describes the data and outlines main facts regarding the career choices of PhDs,
the relevance of these choices to S&E and R&D, and the transition patterns between the
different types of careers.
3.1 Data
The stylized facts in this section are derived from the rarely used Survey of Doctorate Recip-
ients (SDR), collected biennially by the NSF since 1973. This longitudinal survey includes
only individuals who graduated from US PhD programs and resided in the US at the time
of the survey. In the original 1973 survey, the target population included graduates between
1930 and 1973. With every survey a fraction of new graduates was added. Information on
these individuals is obtained from the Doctorate Records File (DRF) maintained by the NSF.
The primary information for the DRF is the Survey of Earned Doctorates, a one-time census
type survey on all doctorates from the US institutions in S&E or health fields collected at
the time of their graduation. Since the latter started in 1957, the information source on the
graduates prior to 1957 are taken from a registry on highly qualified scientists and engineers.
The latter is assembled by the National Academy of Sciences from university catalogues,
federal laboratories, and other sources. This way, in 2001 the survey had a sample of 40,000
individuals representing 650,000 doctorates under the age of 76 residing in the US. Due to
the information collection techniques (multiple follow-ups and phone interviews), the survey
has very high response rates (80 percent). This potential source of error is adjusted in the
survey using the weights. This study bases its analysis on the weighted data. Non-response
8
rates for the employment and personal information used in this study are low (0-1.2 percent
and 0-2.7 percent respectively).
The survey is unique, first of all, because it provides information on a group of profession-
als which is small relative to the population. For this reason other data sets (e.g., CPS) would
have a very small number of individuals with PhD degrees. Second, the data in the SDR
is longitudinal and allows one to follow individuals as their careers unfold. This is crucial
for the analysis of transitions and their duration. Finally, the survey asks about academic
achievements and other profession-specific information usually unavailable in other surveys.
Several studies of doctorates used other data sets, primarily collected by the authors, for
example, Mangematin (2000), Gaughan and Robin (2004), Oyer (2005), Diamond (2001),
Grimes and Register (1997). Their data sets were compiled in most cases from individuals’
CVs posted on their web sites. The advantage of this approach is that it allows one to
construct complete employment histories and have full productivity information unavailable
in the SDR. Unfortunately, I could not use a similar approach or use the data sets from the
studies described above because of the small number of observations which also come from
the academic sector only.
Survey questions related to employment were used in this study to analyze the types of
employment held by doctoral scientists and engineers. They contain information on employ-
ers (e.g., region, sector, size, detailed type of industry), occupations and their relevance to
the degree major according to the NSF taxonomy, primary and secondary activities on the
job including specific activities (e.g., basic and applied research, design, and development),
information on professional activity (e.g., membership in scientific societies and participation
in conferences), and scientific productivity variables (e.g., publications and patents). Ques-
tions about research productivity were collected only in selected years (1983, 1991-2001)
and did not include publications or patents obtained prior to graduation. These questions
had very low response rates. The only variable suitable to serve as a proxy for the quality
of a researcher that was consistent throughout the entire period was the information on
the degree-granting institution (e.g., institution code, its public or private status, and its
9
Carnegie classification).
The employment information collected by the survey was problematic because it was
not retrospective. That is, the individuals were asked only about their last two places of
employment. The response rates regarding the earliest of these two places were low, and
these responses were not consistent. For example, some individuals stated that they changed
jobs, whereas their responses about their previous occupations and employers indicated
no changes. For this reason, only responses regarding the most recent employment were
considered in this study. The longitudinal nature of the survey allowed construction of
the employment histories and task-to-task transitions for the period during which a person
responded to the survey. However, this approach had several problems. First, some histories
had gaps due to non-responses in some years. Second, there were career histories that had
missing information at the beginning because some individuals were included in the survey
years after they had graduated.
The sample was limited to white men with PhDs in natural sciences1 and engineering, who
were employed full-time, had no secondary employment, and responded to all employment-
related questions in at least three consecutive surveys. Due to these conditions, the sample
size was reduced to 15,000 individuals, whose characteristics are described in the last column
of Table 3. Individuals in the estimation sample were on average 45 years of age, US native
citizens (82%), unmarried (53%), with a life science or physical science degree (44% and
28% respectively), and had graduated from public universities (66%) that were most likely
to be a Research I/II institution in the Carnegie classification (88%). They were more likely
to work in the academic sector (49%). Among those in academia, majority was tenured.
Finally, the sample consisted mostly of the cohorts which graduated between 1970 and 1975.
3.2 Types of careers and assignment principles
In this section, data on occupations, primary activities on the job, and employer sector are
analyzed to determine the extent to which the careers of scientists are related to R&D.
1Natural sciences include mathematics and computer sciences, life and health sciences, and physicalsciences.
10
First, I evaluated occupations and their relevance to R&D. The survey respondents were
given a list of occupations and asked to select one that most closely resembled their primary
occupation at the time of the survey (see the Appendix for the list of major occupational
groups). I grouped all occupations into several categories: a) scientists, b) engineers, c) post-
secondary teachers, d)top- and mid-level managers, and e) non-S&E others. A comparison of
the occupational choice by scientific field showed that individuals were concentrated mostly
in occupations defined as “scientist” (“engineer”) or “post-secondary teacher” in the same
discipline as the major of their degree (Table 1), with some minor exceptions. For example,
some mathematicians or physicists by training worked as teachers in engineering or life
sciences programs. For the purposes of my analysis, I chose these aggregate occupational
groups rather than detailed occupations.
To have a better understanding of the relevance of these occupations to R&D, I ana-
lyzed primary activities in each of these occupations. The respondents were asked to report
activities that occupied most of their time on a typical week at their primary job. They
were offered a list of activities containing 14 choices as a reference. These activities could be
grouped into a) R&D activities (e.g., basic and applied research, development and design);
b) non-R&D activities (e.g., teaching, professional services, software design); c) managerial
and administrative activities (e.g., marketing, sales, quality control, general management),
and d) non-S&E activities (e.g., finance and accounting). The activities in the first group can
be thought of as creation of new knowledge, whereas the activities in the second group are
oriented towards application of accumulated knowledge and technologies. The third group
is not as easy to interpret, because it is not stated whether the person directed a research
laboratory, production facility, or a financial firm. The fourth set of activities is clearly
unrelated to S&E activities.
Cross-tabulation of these activities by occupation shows that occupations include a mix-
ture of activities which have varying relevance to R&D (Table 2). In addition, since an
occupation was self-reported, there were cases where major activities did not agree with the
reported occupation. For example, some individuals classified themselves as physicists, but
11
their primary activities were reported as accounting and finance. This could be a physicist
by training employed in an occupation unrelated to S&E (e.g., financial analyst). Another
example is when individuals are recorded as “teachers” but report R&D-type activities and
no teaching at all. It is possible that these individuals were academic scientists and thus
were categorized as “post-secondary teachers” (probably recoded by the NSF) even if they
did primarily research. Therefore, judging the relevance of employment to R&D from occu-
pations alone can be misleading. Moreover, after studying the industry of the employer, it
became evident that some individuals reported themselves as scientists even if they worked
in non-S&E industries, such as financial and other business services. The analysis of the
activities of these scientists showed that they were mostly involved in accounting, finance, or
professional services. The conclusion was to study employment tasks corrected by industry
to infer the relevance of employment to S&E in general and R&D in particular.
Using information on occupations, industry of employment, and primary activities, I
divided all jobs into three types according to the main task they involve. Because information
on primary activities played a crucial role in assigning a job to a particular type, they are
referred to as “tasks” throughout the remainder of the paper. Two of the tasks belong
to the S&E sector. The third task includes jobs unrelated to S&E and is further referred
to as a non-S&E task. I distinguished between S&E and non-S&E tasks to pin down a
“career change” type of mobility as opposed to other types considered in previous studies.
Within S&E, jobs differed in required skills, prices, depreciation rates, and other features.
To reflect this, I further divided them into two groups: R&D and applied tasks. I assigned
individuals to be in the R&D task if their occupation was a scientist, an engineer, or a “post-
secondary teacher” if they indicated R&D as their primary activity and they were employed
in sectors other than financial and business services. I also included individuals in certain
managerial occupations, which involve R&D activities, supervision of R&D activities, or
imply employment in R&D related sectors.
A job was classified as “applied” if individuals reported a) their occupations as “post-
secondary teachers”, “technicians and technologists”, “surveyors”, “sales and marketing spe-
12
cialists”, and “management and administration”, b) indicated applied activities as their pri-
mary activities, and c) their employment sectors belonged to S&E. Individuals in applied
jobs were expected to be involved in mostly teaching S&E subjects, performing professional
services in S&E (e.g., technical consulting, project evaluations, and surveying), or manage-
rial and sales activities in these areas. Note that according to the NSF taxonomy most of
these individuals were recorded as working in non-S&E jobs.
The last category, non-S&E tasks, included tasks unrelated to S&E. In order to distin-
guish the relevance of a job to S&E, I used a taxonomy different from that adopted by the
NSF. For example, the NSF taxonomy considers all managerial occupations as unrelated
to S&E. This way, a person becoming the head of a university department or a director
of a research laboratory would be considered as changing his career. For the purposes of
my analysis, such a classification would give misleading results because mobility from S&E
to non-S&E would include both career advancements (exits due to promotions) and career
changes. Distinguishing between the two is possible only when managerial occupations are
separated by their relevance to S&E. One example of a non-S&E task would be teaching
non-S&E subjects (e.g. in humanities, business2, law, or arts). Another example would be
employment in the areas of legislation, business services such as finance, accounting, non-
technical consulting, or marketing and sales of products and services in non-S&E industries
(e.g., tourism and hospitality, entertainment, and media). While business consulting or
legislation in high-tech industries or manufacturing requires technical knowledge, technical
education for these professions does not require to be at the level of a research doctorate.
3.3 Career choices and empirical transitions
Overall, 56.7 percent of all doctorates in the sample were employed in R&D-related tasks.
Applied tasks accounted for 35.4 percent of all employment. The remaining 7.8 percent
worked in non-S&E tasks. These facts contradict a common perception that the S&E skills
of doctorates are narrowly specialized and find no utilization outside their profession.
2For certain social science majors, especially for economists, business would not be an unrelated area.However, social science doctorates were excluded from the estimation sample.
13
The first three columns in Table 3 report the summary statistics by task. The main
observation from the summary statistics is that R&D tasks are performed by the youngest
(44 years old) and non-S&E tasks by the oldest PhDs (49 years old). Non-S&E tasks were
predominantly found in industry, followed by R&D, and then by applied tasks (75%, 48%,
and 25% respectively). Finally, doctorates employed in R&D tasks were the least likely to
report having a non-S&E degree (8%) compared to 23% in non-S&E tasks. This may sug-
gest that individuals employed in non-S&E tasks obtained non-S&E skills or demonstrated
interest in non-S&E before joining their doctoral training. Doctorates in academic R&D
were less likely to be tenured than the average (44% vs. 51%).
Task choices varied throughout a career. As shown on Figure 1, early in an individual’s
career participation in R&D tasks was high (75 percent), while employment outside S&E was
very low (3 percent). However, only 45 percent of doctorates were still in R&D thirty years
after graduation. Employment in applied tasks grew from 25 percent right after graduation to
42 percent at the end of a career. Non-S&E tasks accounted for 10 percent of all employment
twenty years after graduation and remained stable thereafter. The next section describes
the transition patterns in more detail to better understand their nature.
3.3.1 Mobility patterns
Transitions between different types of tasks can be divided into transitions within S&E
versus transitions out of S&E. For 15,000 individuals in the sample, there were on average
1.73 task-specific spells. Unfortunately, this means that repeated spells of the same type were
not observed, which made it impossible to estimate a duration model with dependent risks
using the joint distribution of unobserved heterogeneity3. The longest observed spells were
28 years long. However, the number of individuals in these spells was fewer than 100 people
in total. For this reason only the spells with more than 20 observations were considered,
which shortened the longest spell to 18 years. The seven most common career paths in the
sample were a) uninterrupted employment in R&D (28%), b) uninterrupted employment in
3See D’Addio and Rosholm (2002) for a discussion of identification problems in this type of models
14
applied tasks (22%), c) careers starting in R&D and eventually ending in application (12%),
d) careers starting in application and eventually ending in R&D (6%), e) employment in
R&D with a short intermediate spell in an applied task (5.35%), f) employment in applied
tasks with a short intermediate spell in R&D (3.28%), g) careers starting in either R&D
or application and eventually ending in non-S&E (2.8 percent and 2 percent respectively).
The remaining 18 percent of career histories had other patterns. Table 4 contains average
transition rates. That is, each number is computed as a fraction of individuals who changed
task i for task j between periods t− 1 and t over all individuals in task i at time t− 1. One
can notice that R&D-related tasks have slightly higher retention rates compared to those in
applied tasks: 0.61 versus 0.58 respectively. In addition, average transition rates to non-S&E
were similar: 0.07 out of R&D compared to 0.08 out of application. The next subsection
considers the transition dynamics in more detail.
Reallocations within S&E
Figures 2 and 3 demonstrate transition rates between S&E tasks by discipline over time.
The horizontal axis depicts task-specific tenure at the time of the transition. The empirical
exit rate eij(t) on the vertical axis is the fraction of individuals who left task i for task j
after being employed in task i for t years out of the total number of the employed in task i
for t years. There are several interesting observations that are worth noting. First of all, the
transition rates from R&D to application are non-monotonic: They are stable within the first
10 years of employment, then they fall by half by year 14, and increase thereafter. This type
of transition occurs in two “waves”: the first one within the first 8 to 10 years in occupation
and the second after 14 to 18 years. Notably, the first wave occurs during the time believed
to be crucial for building the foundation for reputation and recognition in R&D. The second
wave could include established scientists who leave R&D for consulting and similar tasks
(Zucker et al. (1997), Cater and Lew (2005)). This is consistent with the predictions of
the human capital theory. The transitions are notably high for mathematicians (twice of
that in other disciplines), especially within the first 14 years. This might suggest higher
competition for R&D positions in mathematics, especially considering that most of these
15
people are employed in academic institutions with tenure requirements, or higher return on
experience in R&D for mathematicians in applied tasks.
Mobility out of applied occupations back to R&D monotonically decreases within 12-
14 years of employment, which might suggest that R&D related skills depreciate quickly.
Different disciplines experience different transition rates although very similar in pattern and
timing. They peak at 4 years of task-specific experience from as high as 0.14 for engineers to
as low as 0.1 for mathematicians. The transition rates fall to 0.06 and 0.02 respectively by
14 years of experience. The observed difference in transition rates by discipline can suggest
different depreciation rates of R&D skills or a different return in R&D on experience in
applied tasks. But overall, the timing of the transition rates to R&D is consistent with a
belief that returns or late entries to R&D are possible only within a short period of time.
Transitions out of S&E
Figures 4 and 5 show transition rates from R&D and application to non-S&E. The first
observation is that in both cases the rates are much lower than those within S&E. The
second is that the transition rates are non-monotonic: The rates out of R&D are mostly
single-peaked and similar in shape for all disciplines, while those out of applied occupations
differ substantially by discipline in shape and magnitude. Transitions out of applied tasks
are lower than those out of R&D and do not exhibit a simple pattern. The third observation
is that the transitions out of R&D die out after 16 years of R&D-specific tenure and are
at their highest between 8 and 14 years in R&D, which corresponds to the period believed
to be crucial for building a foundation for the development of scientific reputation. The
peaks at 8, 10, or 12 years may correspond to several postdoctoral appointments, which
are on average 2-3 years long, or a combination of postdoctoral appointments and a tenure-
track appointment, which is on average 6 years long. Some studies find that postdoctoral
appointments became more common and much longer in the last decade or so, and in some
fields they became a prerequisite for a tenure-track position (Stephan and Ma (2004), and
Robin and Cahuzac (2003)). It is possible that individuals who leave R&D for non-S&E
occupations are those unable to land a tenure-track position. To test this hypothesis more
16
explicitly, I controlled for the type of employment and number of postdoctoral appointments
when estimating a transition model described in more detail in the next subsection.
4 Model
In order to understand the reasons and the timing of the transitions, I developed an econo-
metric duration model of transitions and estimated it for four types of transitions: 1) from
R&D to applied tasks, 2) from applied to R&D tasks, 3) from R&D to non-S&E tasks, and
4) from applied to non-S&E tasks.
4.1 Specification of the transition model
In every period individuals can be in one of three states denoted as s(t), corresponding to
“R&D”, “application”, and “non-S&E”, denoted respectively as 1, 2, and 3. Let the rate of
each transition be denoted by:
µij(t∣X) = lim∂t→0
P [s(t+ ∂t) = j∣s(t) = i]
∂t, i, j = 1, 2, 3, i ∕= j
where X is a set of observable characteristics. Once state i is chosen, the duration of the stay
in i is determined by µi =∑
i µij, i ∕=j. Regardless of the initial state, an individual can exit
into one of the two remaining states. Suppose, there exist 3 latent duration variables, {Tj}3j=1,
with Tj corresponding to the length of stay in the initial state i before exiting to state j. In
terms of the terminology of the duration analysis, they correspond to two competing risks
that affect the length of stay in the initial state. In empirical duration analysis literature,
it is often assumed that competing risks are independent which is much easier to deal with
computationally. In most economic problems, this assumption is often very difficult to justify
because individual characteristics (e.g., skills or preferences for different sectors) are more
likely to be correlated. This dependency can be captured through the joint distribution
of some unobservable characteristic " over the three states, that is, transition rates are
independent only conditional on ". For computational reasons, "’s are often assumed to be
discrete with a finite support and a joint probability mass function G("). The identification
17
of these types of models requires data sets with multiple spells of the same type. However,
the data set used in this study contains at most one spell of each type, and, therefore, does
not contain enough information to include unobserved heterogeneity. Therefore, I follow the
traditional approach and assume that the risks are independent conditional on the choice of
the observable characteristics.
The set of observable characteristics X includes four groups: 1) demographic and educa-
tional characteristicsDEM , 2) job and activity characteristics JOB, 3) research productivity
variables PROD, and 4) a set of macroeconomic variables MACRO. The first group DEM
includes:
DEM = {AGE,CLASS,MARIND,CITIZEN,MAJOR, TOPSCHOOL,NSE}
The first two variables, AGE and CLASS, are two sets of indicators for different age groups
and graduation years. They control for life cycle and vintage effects on the duration of
employment. Grouping into different age and vintage categories is done to preserve the
proportionality of the hazard ratio over time. Next, MARIND indicates that a person is
married or in a common-law relationship. CITIZEN is an indicator for native or naturalized
US citizens. MAJOR is a set of indicators for the field of PhD. TOPSCHOOL is an
indicator for graduates from both Carnegie Research I/II and highly-ranked schools4. Finally,
NSE is an indicator that individuals had a BA or MA in non-S&E fields5 prior to their PhD.
This indicator is expected to control for non-S&E-specific skills or preference to non-S&E
tasks.
The next set of characteristics, JOB, describes major activities and features of the task,
from which the transition originated:
JOB = {SECTOR, TENURE,POSTDOC, 2PDOCS,CARNRES,ACT}
where SECTOR contains an indicator for the employment sector (academia, government,
or industry), TENURE contains indicators for tenure status if individuals were employed
4These include CalTech, UC Berkeley, Stanford, MIT, Harvard, Princeton and Yale.5Social sciences, arts, humanities, business or law
18
in academia, POSTDOC is an indicator for a postdoctoral appointment, 2PDOCS indi-
cates that a person completed more than two postdoctoral fellowships, CARNRES indicates
that if the employer is an academic institution it belongs to a Carnegie Research I/II cat-
egory, and finally ACT indicates primary activities on the job and serves as a proxy for
occupation-specific skills. Two variables are also used to control for the effect of postdoc-
toral appointments: an indicator that a current position is a postdoc (POSTDOC) and
an indicator that a person held two or more postdoctoral appointments (2PDOCS). This
last indicator is used instead of the number of appointments in order to preserve hazard
proportionality.
The third set of characteristics includes several possible measures of research success:
PROD = {ARTGRAD, ART, PAPERS, PATENTS}
where ARTGRAD and ART contain the number of publications (at graduation and in total,
respectively), PAPERS contains the number of working papers in two years preceding the
survey, PATENTS includes the total number of granted patents. This set of variables is
expected to capture the effect of research productivity on transitions.
The final group, MACRO, contains the following variables:
MACRO = {RD,UE,ENROLL}
where RD is the share of R&D expenditures in the GDP, UE is the unemployment rate,
ENROLL contains the enrollment rates in undergraduate programs in S&E. All other vari-
ables in MACRO correspond to the year preceding the survey. The R&D rates are a proxy
for the demand for research skill, ENROLL is a proxy for the demand for teachers, which
represents a part of those employed in applied occupations, and, finally, UE represents over-
all economic conditions. The descriptive statistics for the variables can be found in Table
3.
19
4.2 Functional forms
Each transition probability µij is assumed to have the following functional form:
µij(ti∣X) = µ0ij(X) ⋅ µij(ti),
where µij(ti) is a baseline hazard and µ0ij(X) is a systematic hazard. To derive the likelihood
function, we need to consider contribution of spells of different types. For a single spell
k which does not end in an exit to either of the states (i.e., a right-censored spell), the
likelihood contribution is simply the survival function, Sk(t) = 1−Fk(t) = exp[−∑ti=1 µk(i)].
Contribution of the spells ending in either of the states would be equal to:
Lk = S1−∑
j ±kj(t)
k
∏j
µkj(t)±kj(t)
where ±kj(t) is an indicator that an exit from k to j occurs at time t. Since each individual
can have more than one spell, denote R the number of spells that an individual has in R&D,
A in applied occupations, and N in non-S&E. For this individual the likelihood contribution
is:
Li(°;X) =R∏j
Lrj
A∏
k
Lak
N∏
l
Lnl, (1)
where X is the matrix of observable characteristics, and ° is the vector of parameter values.
Only the spells observed from their start were included, that is, no left-censored spells were
used. Therefore, no correction for initial conditions were used.
Finally, the loglikelihood function over all individuals is:
l(°;X) =∑i
ln(Li(°;X)) (2)
The final loglikelihood function (2) is maximized with respect to the parameters on the
observable characteristics ¯ij and baseline hazard parameters µmij .
4.3 Baseline hazard
I assume an exponential distribution of the survival times, which in turn implies an expo-
nential, or constant, baseline hazard. This assumption can be applicable to labor market
20
models if one slightly modifies it by allowing baseline hazards to be constant only on certain
intervals, that is, piecewise constant. The time is divided into 15 periods of 2 years, with
the 15th period being in (28,∞) interval. For the mapping principles between time and the
intervals and modifications of the hazards functions to accommodate that, see D’Addio and
Rosholm (2002). The baseline hazard becomes:
µik(t) = exp(µj(t)ik ), j(t) ∈ (1, 15)
Analysis of hazard proportionality suggested the proportionality assumption was violated for
postdocs. One way to solve this problem is to stratify the model for postdocs by introducing
both a postdoc-specific intercept and a postdoc-specific baseline hazard.
5 Estimation results
Several specifications of the model were estimated. The first specification includes only
personal and employment characteristics, DEM and JOB. The next four specifications
sequentially introduce four different research productivity variables from the PROD set of
characteristics. The sixth specification includes an indicator if the person had more than one
postdoctoral appointment. The final specification includes labor market variables, MACRO.
All seven specifications were estimated on the pooled sample.6 The base categories are
“engineering” for MAJOR, “government” for SECTOR, “under 30” for AGE, and “post-
1990” for CLASS. Base categories for activities are a) design and development in R&D and
b) professional services in application.
Mobility within S&E
Tables 5 and 10 present estimation results for the transitions between R&D and applied
tasks. We are mainly interested in answering the following questions: Is the observed mobility
age- and cohort specific? Does it depend on research productivity? How is mobility affected
by the labor market conditions?
First, specifications without productivity variables are analyzed. The first result is that
6Estimation by discipline produces similar results.
21
there is no significant cohort effect on transitions within S&E except for the class of 1980-
1989. The age effects are significant only for the transitions from application to R&D. The
estimates suggest that older doctorates are less likely to leave applied tasks to return to R&D.
Next, scientists are more likely to leave R&D for applied tasks compared to engineers with
an exception of computer scientists. This suggests that either the competition in scientific
R&D is more intense, or that the skills and knowledge accumulated in scientific research are
more applicable in applied tasks than those in engineering. Mobility from applied tasks back
to R&D exhibits no significant differences by discipline. The estimates show that individuals
engaged in research (both applied and theoretical) and managerial activities were less likely
to leave R&D for application compared to those in design and development. The next result
is that substantial differences exist in transition rates from R&D to application by sector after
one controls for demographic characteristics. The most likely to leave R&D are the doctorates
employed in the academic sector, followed by those in governmental jobs. Notably, the more
advanced the tenure status of academics, the likely they are to leave R&D for applied tasks.
This finding may suggest that academic scientists invest in R&D early in their careers to
enhance their total scientific knowledge, which they later “sell” to their students or clients
depending on the types of applied task they choose, as was suggested by Cater and Lew
(2005). At the same time, transitions to R&D from applied tasks are made mostly out of
non-academic sectors. Among academics those who are tenured are the least likely to switch
from applied tasks to R&D. Finally, individuals employed in Carnegie Research universities
have lower probability of leaving R&D for applied tasks, and higher probability to leave
applied tasks for R&D. In the first specification, the “top school” indicator is the only
measure for quality of the individual’s ability or skills. The estimates show that top school
graduates are more likely to switch to R&D after employment in applied tasks. However,
their probability of leaving R&D for applied tasks is insignificant although also positive. The
results of the benchmark model do not change when the set of variables is augmented with
an indicator of postdoctoral appointments or variables of economic conditions, which are
described later on.
22
Up to this point, only the individual characteristics and job features were considered.
It would be interesting to see, however, whether the transitions within S&E are affected
by general economic conditions. Three variables were considered: 1) unemployment rate
for the entire labor force, 2) the share of the total R&D expenditures in the GDP, and 3)
enrollment in S&E programs in universities and colleges. The estimates suggest that labor
market conditions at the time of a transition matter for mobility within S&E. High R&D
expenditures decrease mobility both out of and into R&D, while higher unemployment rates
increase mobility. Enrollment rates, which represent the demand for labor in teaching, have
a negative but insignificant effect on mobility from R&D to application, and positive and
significant effect on mobility out of the applied tasks to R&D. It is possible that when the
demand for teachers is high, it creates demand for researchers as well since in academia
these two activities are combined. This way, some scientists performing professional services
switch to academic jobs when demand in academia is high although they spend more time
doing research than teaching, therefore, switching tasks. This is one of the examples when
switching tasks is accompanied by changing jobs.
Mobility out of S&E
Tables 7 and 8 present the estimation results for the transitions to non-S&E from R&D
and application respectively. The first finding is that in the basic model there are no cohort
or age effects on mobility, unlike the case with the mobility within S&E. Second, out of all
disciplines only life scientists are more likely to leave R&D for non-S&E. One reason for
that may be a higher competition in R&D positions for life scientists due to the increased
number of students in the field or decreased number of positions (Tilghman et al. (1998)).
Alternatively, the R&D skills of life scientists are more valued outside S&E than those of
other majors. As for the mobility from applied tasks to non-S&E, there is no significant
difference between scientists and engineers. The only exception are computer scientists who
are the least likely to make the switch. Next, individuals with non-S&E degrees are more
likely to switch to non-S&E tasks regardless of the origin of the transition suggesting that
these individuals possess the skills required in non-S&E tasks or have preferences towards
23
non-S&E. Academics are the least likely to leave S&E. This might suggest that employment
in the industry or government makes the individuals more exposed to non-S&E employers,
or that these sectors provide skills valued outside S&E.
Finally, labor market conditions also seem to have significant effects on the mobility out of
S&E. Just as the case with the mobility within S&E, the transitions out of S&E seem to slow
down during the periods of high R&D expenditures, with the effect being more significant
for exits out of R&D. Enrollment rates which were proxies for demand for post-secondary
teachers and maybe academic researchers with teaching duties have the same significant
negative effect. Finally, mobility is higher during the periods with higher unemployment
rates. This can be explained by budget cuts for the academic sector which accounts for half
of the researchers. The budget cuts can affect hiring and dismissal policies, structure of the
department with emphasis on temporary faculty, and decrease in funding opportunities and
acquisition of new laboratory equipment and materials.
Analysis of the baseline hazards shows that most of the time dummies have significant
coefficients, thus exhibiting effects not picked up by the chosen explanatory variables. This
is also true for the time dummies interacted with the postdoctoral status. If we add labor
market conditions as controls, then some of the time dummies become insignificant, but
others remain significant. Estimation on the limited sub-sample with non-missing responses
to the productivity variables produces less significant effects of the baseline hazard, which
may be a result of the effects of the productivity variable or problems with the sample
selection described earlier. It remains a task to explain the variation in transition rates in
most periods.
Research productivity and mobility
Research productivity is an important indicator of success in R&D. Four measures of
productivity were considered in this study, but only the estimates for the total number of
published articles are shown as representative of the remaining three. Response rates for all
questions regarding productivity were very low. The subsample with non-missing responses
to the productivity questions is not only smaller but it also differs from the main sample
24
for each set of transitions. For example, the failure rates in the main sample are much
higher than those in the subsample with productivity variables, which indicates a highly
selective group of doctorates less prone to leave S&E. Therefore, the estimation results for
this subsample should be interpreted with great caution.
According to the estimates, individuals with a higher number of publications do exhibit
lower transition rates from R&D to application, but this effect is insignificant. However,
individuals who published more were significantly more likely to switch from applied tasks
to R&D suggesting that returns or late entries to R&D are conditional on active publishing
or patenting activity even if R&D is not a primary activity. Even though this is commonly
perceived to be true, so far no-one has empirically evaluated it. In this specification the
magnitude, sign, and significance of some variables changes compared to the baseline model,
most likely due to a smaller and non-random subsample. For example, the cohort effects
become significant at 1 percent, meaning that the probability to leave R&D is higher for
younger cohorts after controlling for the age effects. In this specification significance at
tenure track status disappears for non-tenured academics together with the cohort effects.
For transitions from S&E to non-S&E tasks, the results suggest that individuals with a
higher number of publications are less likely to leave both S&E tasks for non-S&E, but in
both cases the coefficients are insignificant. As was the case with the mobility within S&E,
the indicators of tenure status in the academic sector lose their significance in this subsample
except for the postdocs in R&D and tenured academics in application with both groups being
less likely to leave S&E. At the same time, the indicators of graduation period, or vintage,
gain significance for transitions from application to R&D. Finally, other coefficients change
their magnitude but not their signs in this specification.
Postdoctoral appointments
An analysis of the transitions in and out of R&D (and S&E in general) allows to better
understand the nature of postdoctoral appointments. There are several competing expla-
nations about the nature of and reasons for postdoctoral appointments. One explanation
is that research methods become more sophisticated or equipment required for research be-
25
comes more expensive. Therefore, researchers need to go through additional training before
they can handle that equipment independently and this training is provided during post-
doctoral appointments. In this case the probability of surviving in R&D should increase
with the number of postdoctoral appointments. Alternatively, postdoctoral appointment is
a probationary position required to observe potential candidates for a job and eliminate un-
suitable ones, for example, to reduce the risk of damaging expensive laboratory equipment.
In this case, the more postdoctoral appointments individuals have, the more likely they are
to eventually leave S&E altogether. Another possibility is that postdoctoral appointments
are results of the oversupply of scientists, and a perfectly competitive market reacts by re-
ducing wages and, thus, creating postdoctoral positions. However, this explanation does not
have a direct implication for mobility between tasks and cannot be tested with the approach
used in this paper.
The results show that having more than one postdoctoral appointment in the past had
a negative but insignificant effect on transitions from R&D to application. However, having
more than two postdoctoral appointments significantly reduced the chances of going from
application back to R&D. An interesting result is obtained for the transition from R&D to
non-S&E tasks. The effect of multiple postdoctoral appointments is positive and significant
for these transitions. This suggests that the longer it takes a person to secure a position
in R&D, the more likely the person is to leave S&E altogether, which can resemble the
“discouraged worker” effect for the exits from unemployment spells. This makes postdoctoral
appointments “waiting lists” or probations rather than skill-enhancing training. If the latter
was the case, then it would manifest itself in a negative significant effect on transitions out of
R&D as it would make individuals who had more training to be more suitable and successful
in R&D and prevent the transitions.
Baseline hazard
Baseline hazard in duration analysis is the hazard when all used covariates are equal to
zero. Therefore, it picks up the part of transitions unexplained by the used covariates. In
this case, the baseline hazard was split into several periods, and therefore the coefficients
26
at baseline hazard in each period show how much mobility remained unexplained. From
Tables 9, 10, 11, and 12, we can see that in all specifications for all types of transitions
most coefficients at the piece-wise constant baseline hazards remain highly significant. This
means that more work needs to be done to explain the dynamic pattern of task transitions.
A potentially fruitful extension could be to introduce unobserved heterogeneity that would
generate correlation between the risks, which was not feasible in the data used in this study.
6 Conclusion
This paper proposes a new approach to study careers of doctorates in natural sciences and
engineering. It conducts a careful empirical analysis of the dynamics of doctoral careers
in S&E in general and R&D in particular. In this study career dynamics were described
as transitions between different tasks and their relevance to R&D, application of scientific
knowledge, and non-S&E. Finally, it was empirically evaluated how research productivity,
labor market conditions, and other covariates affect these career dynamics.
The first finding is that only 57 percent of all PhDs worked in R&D tasks. Another
35 percent worked in applied tasks such as teaching, software development, or professional
services. I also found that, contrary to common belief, S&E doctorates find employment
outside S&E. In the data 8 percent of all S&E doctorates held positions outside S&E fields
in sectors such as finance or business services. The distribution of PhDs in all three types of
tasks changed as their careers developed. The majority of doctorates started their careers in
R&D (72 percent), but only 45 percent were still in R&D 30 years later. About 80 percent
of those who left, switched to applied tasks, while the rest moved out of S&E completely.
Most transitions occurred during the time believed to be critical for researchers to build
a reputation in their field. An analysis shows that it was mostly academic researchers on
tenure-track or tenured who switched to the applied tasks, which agrees with the predictions
of the human capital theory. Returns from applied to R&D tasks were possible with the
evidence of high research productivity. It was also found that individuals who leave S&E
are more likely to be employed in non-academic sectors and have non-S&E degrees prior
27
to their PhD. The probability of leaving R&D increases with the number of postdoctoral
appointments, which was interpreted as evidence that postdoctoral appointments serve as
waiting lists for permanent academic appointments rather than provide additional training.
Analysis of the labor market conditions at the time of the transition shows that mobility
both within and out of S&E is sensitive to both general economic conditions measured
by unemployment rates, and conditions specific to S&E measured by the share of R&D
expenditures in GDP and enrollment rates in S&E programs in universities and colleges.
Analysis of the baseline hazard shows in all specifications it remains a non-constant function
of time periods since graduation because the covariates were not able to completely explain
the career dynamics.
The results are new and interesting, but it remains important to understand and model
the mechanisms that generate these outcomes in order to create effective policies to attract
and retain the best doctoral scientists in R&D.
References
Almeida, P. and Kogut, B. (1999). Localization of knowledge and the mobility of engineers
in regional networks. Management Science, 45(7):905–917.
Audretsch, D. and Stephan, P. (1999). Innovation, Industry Evolution, and Employment,
chapter How and why does knowledge spill over?, pages 216–229. Cambridge University
Press.
Biddle, J. and Roberts, K. (1994). Private sector scientists and engineers and the transition
to management. The Journal of Human Resources, 29(1):82–107.
Cater, B. and Lew, B. (2005). Theory of tenure-track contracts.
D’Addio, A. and Rosholm, M. (2002). Labor market transitions of French youth. University
of Aarhus, Working Paper.
Diamond, A. (2001). Scientists’ salaries and the implicit contracts theory of labour markets.
Int.J. Technology Management, 22(7/8):688–697.
Fallick, B., Fleischmann, C., and Rebitzer, J. (2005). Job hopping in Silicon Valley: Some
evidence concerning the micro-foundations of a high technology cluster. NBER, Working
Paper(11710).
Ferrall, C. (1997). Empirical analysis of occupational hierarchies. The Journal of Human
Resources, 32(1):1–34.
Gaughan, M. and Robin, S. (2004). National science training policy and early scientific
careers in France and the United States. Research Policy, 33:569–581.
Grimes, P. and Register, C. (1997). Career publications and academic job rank: evidence
from the class of 1968. The Journal of Economic Education, 28(1):82–92.
Levin, S. and Stephan, P. (1991). Research productivity over the life cycle: Evidence for
academic scientists. The American Economic Review, 81(1):114–132.
29
Mangematin, V. (2000). PhD job market: professional trajectories and incentives during the
PhD. Research Policy, 29:741–756.
Moen, J. (2001). Is mobility of technical personnel a source of R&D spillovers? NBER,
Working Paper(7834).
Oyer, P. (2005). The macro-foundations of microeconomics: Initial labor market conditions
and long-term outcomes for economists and MBAs. Working Paper, Stanford Business
School.
Robin, S. (2002). The effect of supervision on the Ph.D. duration, publications, and job
outcomes. UCL, IRES, Working paper(2002041).
Robin, S. and Cahuzac, E. (2003). Knocking on academia’s doors: An inquiry into the early
careers of doctors in life sciences. CEIS, 17(1):1–23.
Sauer, R. (1998). Job mobility and the market for lawyers. The Journal of Political Economy,
106(1):147–171.
Scafidi, B., Sjoquist, D., and Stinebrickner, T. (2007). Do teachers really leave for higher
paying jobs in alternative occupations? Advances in Economic Policy and Analysis,
forthcoming.
Stephan, P. (1996). Economics of science. Journal of Economic Literature, 34(3):1199–1235.
Stephan, P. and Ma, J. (2004). The increased frequency and duration of the postdoctorate
career stage. Georgia State University, Working paper.
Stinebrickner, T. (1998). An empirical investigation of teacher attrition. Economics of
Education Review, 17(2):127–136.
Stinebrickner, T. (2002). An analysis of occupational change and departure from the labor
force. Evidence of the reasons that teachers quit. Journal of Human Resources, 37(1):192–
216.
30
Tilghman, S., Astin, H., and Brinkley, W. (1998). Trends in the early careers of life scientists.Committee on Dimensions, Causes, and Implications of Recent Trends in the Careers of LifeScientists, National Research Council. Molecular Biology of the Cell, 9:30073015.
Van den Berg, G., Holm, A., and van Ours, J. (2002). Do stepping-stone jobs exist? Early careerpaths in the medical profession. Population Economics, 15:647–665.
Zucker, L. and Darby, M. (2006). Movement of star scientists and engineers and high-tech firmentry. NBER, Working Paper(12172).
Zucker, L., Darby, M., and Torero, M. (1997). Labour mobility from academe to commerce. NBER,Working Paper(6050).
Occupation Field of the PhDComp and Math Life sci Phys sci Engineering
Comp and Math Scientists 0.2603 0.0033 0.0259 0.0501Comp and Math Teach. 0.4941 0.0093 0.0075 0.0144Life Scientists 0.0029 0.3704 0.0422 0.0070Life Teachers 0.0000 0.2167 0.0126 0.0033Physicists and astronomers 0.0068 0.0184 0.3971 0.0271Physics Teachers 0.0010 0.0124 0.2034 0.0041Engineers 0.0254 0.0082 0.0572 0.4465Engineering Teachers 0.0078 0.0020 0.0112 0.1851Social scientists 0.0000 0.0038 0.0009 0.0004Social sci Teachers 0.0049 0.0024 0.0006 0.0012Managers 0.1360 0.1823 0.1876 0.2089Non-S&E Teachers 0.0157 0.0606 0.0092 0.0123Other Non-S&E 0.0450 0.1099 0.0445 0.0398
Table 1: Occupational choice in S&E by discipline
Occupation Primary activityR&D Teaching, consulting, prof.svcs. Management and finance
Scientists 0.8412 0.0536 0.0778Engineers 0.7079 0.1207 0.1350S&E teachers 0.2938 0.6663 0.0319Managers 0.2057 0.0908 0.6642Non-S&E others 0.1791 0.1537 0.5427
Table 2: Primary activities of PhDs by occupation
TaskR&D applied non-S&E Overall
Age 44 47 49 45Married 0.471 0.452 0.670 0.470
Citizen, native 0.81 0.84 0.82 0.82Citizen, naturalized 0.09 0.09 0.13 0.10Permanent resident 0.07 0.05 0.04 0.06Temporary resident 0.02 0.01 0.01 0.02
Comp.sciences 0.02 0.03 0.01 0.02Mathematics 0.05 0.12 0.05 0.07
Health/Life sciences 0.45 0.44 0.44 0.44Physics 0.30 0.25 0.28 0.28
Engineering 0.19 0.17 0.22 0.18Has a non-S&E degree 0.08 0.15 0.23 0.12
Class 1990-2000 0.17 0.13 0.13 0.15Class 1980-1990 0.26 0.22 0.24 0.24Class 1970-1975 0.49 0.52 0.51 0.51Class 1960-1965 0.07 0.12 0.13 0.10
Public university graduate 0.65 0.68 0.68 0.66Graduate of Research I and II 0.88 0.87 0.87 0.88
Academic sector 0.38 0.70 0.17 0.49Government sector 0.14 0.05 0.11 0.10
Industry 0.48 0.25 0.72 0.41Acad. x ten.track 0.062 0.075 0.019 0.067Acad. x Tenured 0.168 0.513 0.091 0.250
Acad. x non-ten.track 0.140 0.127 0.078 0.173Has a postdoc 0.116 0.020 0.006 0.074
Unemployment rate 0.064a
R&D in GDP 0.025b
Enrollment, thousands 12,957c
Number spells 9,978 7,574 2,363Number of individuals 14,988
aSource: US Bureau of Labor Statistics.bSource: National Science Foundation/Division of Science Resources Statistics.cSource: U.S. Department of Education, National Center for Education Statistics, IPEDS
Table 3: Summary statistics
DestinationOrigin R&D Application Non-S&ER&D 0.6130 0.3129 0.0741Application 0.3398 0.5801 0.0801Non-S&E 0.2248 0.2461 0.5292
Table 4: Average transition rates.
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
yrs since graduation
fraction
research application non-S&E
Figure 1: Participation rates in three type of tasks, by years since graduation
0
0.04
0.08
0.12
0.16
0.2
2 4 6 8 10 12 14 16 18 20
yrs on task
hazard
math sciences life sciences physical sciences engineering
Figure 2: Transitions from R&D to application, by discipline
0
0.04
0.08
0.12
0.16
2 4 6 8 10 12 14
yrs on task
hazard
math sciences life sciences physical sciences engineering
Figure 3: Transitions from application to R&D, by discipline
0
0.01
0.02
0.03
0.04
0.05
0.06
2 4 6 8 10 12 14 16yrs on task
hazard
math sciences life sciences physical sciences engineering
Figure 4: Transitions from R&D to non-S&E, by discipline
0
0.02
0.04
2 4 6 8 10 12 14 16
yrs on task
hazard
math sciences life sciences physical sciences engineering
Figure 5: Transitions from application to non-S&E, by discipline
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
married
0.302
0.047
***
0.049
0.107
0.300
0.047
***
0.046
0.058
computersci
0.058
0.107
0.277
0.239
0.061
0.107
0.186
0.106
*mathsci
0.487
0.074
***
0.644
0.182
***
0.486
0.074
***
0.505
0.074
***
life
sci
0.287
0.056
***
0.498
0.145
***
0.280
0.056
***
0.311
0.055
***
physicalsci
0.161
0.057
***
0.245
0.151
0.155
0.057
***
0.186
0.056
***
30-40y
rsold
-0.083
0.072
-0.215
0.155
-0.082
0.072
-0.059
0.073
40-50yrs
old
0.065
0.080
-0.137
0.174
0.069
0.080
0.111
0.083
50-65yrs
old
0.123
0.095
-0.233
0.234
0.127
0.095
0.214
0.099
**class
pre-1950
-0.112
0.142
-0.971
0.262
***
-0.114
0.143
-0.026
0.143
class1950-1969
0.076
0.149
-1.001
0.301
***
0.073
0.149
0.308
0.156
**class
1970-1979
-0.096
0.155
-0.822
0.323
**-0.099
0.156
0.238
0.168
class
1980-1989
0.302
0.174
*-0.761
0.360
**0.294
0.174
*0.649
0.195
***
activity-applied
research
-3.522
0.145
***
-3.470
0.248
***
-3.520
0.145
***
-3.614
0.145
***
activity-basicresearch
-3.975
0.146
***
-4.652
0.335
***
-3.978
0.146
***
-4.086
0.146
***
activity-managem
ent
-0.247
0.039
***
-0.976
0.103
***
-0.247
0.039
***
-0.335
0.041
***
topschool
0.011
0.045
0.146
0.111
0.009
0.045
0.009
0.044
non
-S&E
degree
0.069
0.077
-0.105
0.171
0.070
0.077
0.019
0.077
citizen
0.003
0.099
-0.048
0.252
0.001
0.099
-0.020
0.098
industry
-0.146
0.066
**-0.405
0.162
**-0.147
0.066
**
-0.174
0.066
***
acadxten.track
1.234
0.080
***
1.131
0.175
***
1.226
0.080
***
1.300
0.080
***
acad
xtenured
1.267
0.073
***
1.235
0.170
***
1.265
0.073
***
1.271
0.072
***
acad
xnon
-tenure
track
0.940
0.074
***
0.689
0.165
***
0.937
0.074
***
0.944
0.074
***
acad
xpostdoc
-0.666
0.250
***
-1.544
0.899
*-0.676
0.250
***
-0.524
0.247
**acadxCarnegie
ResearchI/II
-0.373
0.048
***
-0.562
0.114
***
-0.373
0.048
***
-0.374
0.048
***
#articles
-0.003
0.004
>2postdocs
-0.031
0.039
UE
rate
0.160
0.019
***
R&D
inGDP
-1.193
0.134
***
enrollment
-0.004
0.009
constant
-2.803
0.217
***
-0.927
0.462
**-2.768
0.222
***
-2.467
0.435
***
Nindividuals
7488
3390
7488
7488
Nfailures
2497
393
2497
2497
Loglikelihood
-4163.216
-440.939
-4163.010
-4095.249
***-sign
ificantat
1%,**-sign
ificantat5%
,*-sign
ificantat
10%
Tab
le5:
Proportion
alhazardmodel
estimationresults,R&D
toap
plication
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
married
0.211
0.048
***
0.191
0.139
0.194
0.049
***
0.007
0.063
computersci
-0.031
0.097
0.210
0.252
-0.025
0.097
0.045
0.099
math
sci
-0.090
0.073
-0.124
0.229
-0.094
0.073
-0.079
0.074
life
sci
-0.083
0.055
-0.085
0.160
-0.102
0.056
*-0.064
0.056
physicalsci
0.002
0.055
-0.065
0.168
-0.020
0.056
0.033
0.056
30-40yrs
old
-0.384
0.113
***
-0.115
0.286
-0.388
0.114
***
-0.408
0.111
***
40-50y
rsold
-0.631
0.117
***
-0.213
0.304
-0.631
0.118
***
-0.657
0.117
***
50-65yrs
old
-0.794
0.126
***
-0.293
0.325
-0.787
0.127
***
-0.767
0.129
***
classpre-1950
-0.042
0.154
-0.633
0.471
-0.055
0.154
0.110
0.157
class1950-1969
-0.122
0.159
-0.536
0.478
-0.132
0.159
0.148
0.168
class1970-1979
-0.104
0.166
-0.701
0.503
-0.116
0.166
0.171
0.181
class1980-1989
-0.089
0.188
-0.349
0.528
-0.108
0.188
0.168
0.207
activity-teaching
-5.196
0.278
***
-4.107
0.405
***
-5.192
0.278
***
-5.238
0.278
***
activity-managem
ent
0.361
0.051
***
-0.270
0.161
*0.360
0.051
***
0.301
0.052
***
topschoolgrad
0.122
0.046
***
0.187
0.139
0.117
0.046
**0.115
0.046
**
non
-S&E
degree
-0.360
0.091
***
-0.332
0.210
-0.350
0.090
***
-0.380
0.090
***
citizen
-0.093
0.107
-0.318
0.314
-0.100
0.107
-0.144
0.107
sector-industry
-0.147
0.062
**-0.303
0.194
-0.151
0.062
**-0.153
0.062
**
acadxtenure
track
-0.317
0.094
***
-0.347
0.249
-0.339
0.094
***
-0.198
0.095
**
acadxtenured
-0.355
0.073
***
-0.445
0.195
**-0.364
0.073
***
-0.332
0.073
***
acadxnon
tenure
track
-0.340
0.084
***
-0.314
0.230
-0.353
0.084
***
-0.337
0.084
***
acadxCarnegie
ResearchI/II
0.705
0.053
***
0.710
0.137
***
0.704
0.053
***
0.687
0.053
***
#totalarticles
0.008
0.004
*>2postdocs
-0.129
0.044
***
UE
rate
0.098
0.018
***
R&D
inGDP
-1.035
0.129
***
enrollmentlevel
0.032
0.009
***
Nindividuals
5937
1962
5937
5937
Nfailures
2214
263
2214
2214
Log
likelihood
-3683.063
-376.486
-3680.237
-3638.120
Tab
le6:
Proportion
alhazardmodel
estimationresults,ap
plication
toR&D.
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
married
1.076
0.114
***
-0.274
0.352
1.089
0.115
***
0.111
0.128
computersci
-0.143
0.291
0.263
0.766
-0.158
0.290
-0.108
0.289
mathsci
0.247
0.215
0.653
0.832
0.247
0.215
0.183
0.218
life
sci
0.433
0.112
***
0.617
0.424
0.499
0.115
***
0.421
0.113
***
physicalsci
0.136
0.110
0.252
0.424
0.182
0.112
0.121
0.110
30-40yrs
old
-0.135
0.146
-0.027
0.581
-0.140
0.147
-0.170
0.145
40-50y
rsold
-0.083
0.164
-0.067
0.660
-0.115
0.165
-0.190
0.167
50-65yrs
old
0.090
0.205
1.274
0.745
*0.055
0.206
-0.018
0.209
classpre-1950
-0.337
0.347
-2.261
1.050
**-0.329
0.347
-0.644
0.355
*class1950-1969
-0.274
0.360
-1.075
1.201
-0.257
0.360
-0.789
0.381
**
class1970-1979
-0.331
0.378
-0.461
1.214
-0.316
0.378
-0.901
0.406
**
class1980-1989
-0.504
0.427
-0.292
1.362
-0.453
0.426
-1.133
0.470
**
activity-applied
research
-1.770
0.131
***
-2.795
0.486
***
-1.770
0.131
***
-1.875
0.130
***
activity-basicresearch
-2.761
0.245
***
-2.769
0.573
***
-2.727
0.245
***
-2.872
0.249
***
activity-managem
ent
0.089
0.082
-0.235
0.304
0.087
0.082
0.032
0.082
topschool
0.105
0.099
-0.522
0.483
0.123
0.099
0.109
0.099
non
-S&E
degree
0.549
0.130
***
0.599
0.374
0.540
0.130
***
0.444
0.123
***
citizen
0.332
0.346
16.017
0.299
***
0.346
0.346
0.095
0.349
sector-industry
0.174
0.108
0.057
0.375
0.183
0.109
*0.141
0.108
acadxtenure
track
-1.205
0.494
**-0.804
0.826
-1.180
0.494
**-1.057
0.495
**
acadxtenured
-0.696
0.233
***
-0.695
0.671
-0.691
0.233
***
-0.629
0.235
***
acadxnon
tenure
track
-0.589
0.230
**-0.765
0.708
-0.569
0.231
**-0.537
0.230
**
acadxpostdoc
-1.370
1.017
-16.471
0.453
***
-1.300
1.016
-1.233
1.022
acadxCarnegie
ResearchI/II
0.125
0.205
0.440
0.482
0.137
0.206
0.069
0.209
#totalarticles
-0.005
0.013
>2postdocs
0.255
0.093
***
UE
rate
0.301
0.046
***
R&D
inGDP
-0.837
0.357
**
enrollmentlevel
-0.116
0.025
***
Nindividuals
7134
3187
7134
7134
Nfailures
677
58677
677
Log
likelihood
-1813.027
-170.596
-1809.425
-1714.484
Tab
le7:
Proportion
alhazardmodel
estimationresults,R&D
tonon
-S&E
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
married
0.504
0.097
***
0.444
0.428
0.517
0.098
***
0.007
0.119
computersci
-1.101
0.366
***
-15.222
0.457
***
-1.104
0.366
***
-1.118
0.361
***
mathsci
-0.287
0.177
-0.991
0.762
-0.285
0.177
-0.352
0.178
**
life
sci
-0.030
0.119
0.034
0.396
-0.014
0.120
-0.088
0.118
30-40y
rsold
0.241
0.325
-0.114
1.341
0.254
0.326
0.146
0.328
40-50yrs
old
0.421
0.326
0.001
1.341
0.430
0.327
0.280
0.332
50-65
yrs
old
0.386
0.335
0.367
1.370
0.390
0.335
0.315
0.342
classpre-1950
0.040
0.338
14.364
3.701
***
0.047
0.338
-0.073
0.345
class1950-1969
0.081
0.341
15.124
2.765
***
0.088
0.342
-0.091
0.357
class1970-1979
0.159
0.353
15.822
4.226
***
0.164
0.354
-0.024
0.374
class1980-1989
0.049
0.406
15.873
3.506
***
0.054
0.407
-0.061
0.438
activity-teaching
-1.741
0.172
***
-1.463
0.550
***
-1.741
0.172
***
-1.716
0.170
***
activity-prof.servics
-1.851
0.210
***
-1.923
0.636
***
-1.851
0.210
***
-1.884
0.210
***
activity-managem
ent
0.521
0.096
***
0.145
0.361
0.520
0.096
***
0.452
0.093
***
topschoolgrad
0.093
0.103
0.246
0.444
0.098
0.103
0.107
0.101
non
-S&E
degree
0.205
0.124
*0.924
0.391
**0.196
0.124
0.174
0.119
citizen
0.520
0.306
*-1.039
1.034
0.527
0.306
*0.465
0.312
sector-industry
0.436
0.123
***
1.184
0.600
**0.442
0.123
***
0.402
0.121
***
acadxtenure
track
-0.837
0.265
***
-0.586
0.890
-0.819
0.266
***
-0.788
0.267
***
acadxtenured
-1.183
0.175
***
-1.475
0.682
**-1.177
0.176
***
-1.125
0.176
***
acadxnon
tenure
track
-0.572
0.189
***
-0.873
0.734
-0.563
0.189
***
-0.506
0.186
***
acadxCarnegie
ResearchI/II
0.629
0.140
***
1.504
0.380
***
0.635
0.140
***
0.597
0.140
***
#totalarticles
-0.008
0.025
>2postdocs
0.104
0.101
UE
rate
0.161
0.042
***
R&D
inGDP
-0.510
0.300
*enrollmentlevel
-0.115
0.025
***
Nindividuals
5477
1797
5477
5477
Nfailures
689
45689
689
Log
likelihood
-1847.037
-130.472
-1846.531
-1803.537
Tab
le8:
Proportion
alhazardmodel
estimationresults,ap
plication
tonon
-S&E.
Variab
leCoeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
t2-t4
-0.250
0.059
***
0.106
0.155
-0.251
0.059
***
-0.073
0.064
t4-t6
-0.332
0.063
***
-0.015
0.199
-0.333
0.063
***
-0.073
0.068
t6-t8
-0.141
0.063
**-0.317
0.142
**-0.143
0.063
**0.032
0.069
t8-t10
-0.218
0.068
***
-0.527
0.163
***
-0.220
0.068
***
-0.123
0.080
t10-t12
-0.541
0.080
***
-0.803
0.287
***
-0.543
0.080
***
-0.294
0.089
***
t12-t14
-0.831
0.095
***
-0.568
0.220
***
-0.834
0.095
***
-0.549
0.105
***
t14-t16
-0.332
0.085
***
-0.434
0.224
*-0.336
0.085
***
-0.107
0.101
t16-t18
-0.387
0.096
***
-0.378
0.246
-0.391
0.097
***
-0.211
0.113
*t18-t20
-0.418
0.127
***
-0.344
0.223
-0.424
0.127
***
-0.237
0.131
*t20-t22
-0.636
0.155
***
-0.471
0.288
-0.640
0.155
***
-0.351
0.164
**
t22-t24
-0.250
0.138
*-0.155
0.358
-0.253
0.138
*0.138
0.151
t24-t26
-0.327
0.187
*-0.307
0.418
-0.329
0.187
*0.279
0.200
t26-t28
-0.627
0.301
**-0.310
0.298
-0.630
0.300
**0.009
0.310
postdocxt2-t4
-0.216
0.491
-10.902
0.959
***
-0.211
0.491
-0.252
0.487
postdocxt4-t6
-0.428
0.941
-0.427
0.940
-0.524
0.941
postdocxt6-t8
1.002
0.356
***
2.133
1.040
**1.009
0.356
***
0.658
0.353
*postdocxt8-t10
-14.466
0.460
***
-12.069
1.294
***
-12.325
0.460
***
-13.672
0.480
***
postdocxt10-t12
-14.865
0.572
***
-10.002
1.375
***
-12.724
0.574
***
-14.139
0.543
***
postdocxt12-t14
-14.607
0.591
***
-12.476
0.589
***
-13.754
0.572
***
postdocxt14-t16
0.715
0.600
-14.326
1.374
***
0.722
0.606
0.584
0.659
postdocxt16-
t18
1.353
0.977
1.364
0.971
1.080
0.964
postdocxt18-t20
-13.151
0.684
***
-11.016
0.683
***
-12.305
0.689
***
postdocxt20-t22
-13.130
0.671
***
-11.345
1.307
***
-10.979
0.672
***
-12.329
0.667
***
postdocxt22-t24
-17.661
1.049
***
-15.513
1.050
***
-16.955
1.048
***
Tab
le9:
Baselinehazard,R&D
toap
plication
.
41
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
t2-t4
0.147
0.048
***
0.199
0.149
0.144
0.048
***
0.251
0.052
***
t4-t6
-0.111
0.057
*-0.080
0.165
-0.112
0.057
*0.017
0.060
t6-t8
-0.211
0.065
***
-0.471
0.194
**-0.212
0.065
***
-0.128
0.068
*t8
-t10
-0.385
0.079
***
-0.320
0.182
*-0.386
0.078
***
-0.277
0.084
***
t10-t12
-0.642
0.093
***
-0.601
0.317
*-0.640
0.093
***
-0.460
0.098
***
t12-t14
-1.211
0.140
***
-0.784
0.643
-1.210
0.140
***
-1.015
0.142
***
t14-t16
-0.552
0.130
***
-0.502
0.503
-0.554
0.130
***
-0.492
0.137
***
t16-t18
-0.577
0.136
***
-1.142
0.356
***
-0.576
0.136
***
-0.561
0.143
***
t18-t20
-0.283
0.172
0.205
0.525
-0.290
0.171
*-0.245
0.180
t20-t22
-0.695
0.257
***
-0.364
0.340
-0.679
0.257
***
-0.589
0.265
**t22-t24
-0.527
0.281
*-0.581
0.973
-0.511
0.281
*-0.349
0.280
t24-t26
-0.154
0.327
-0.011
0.449
-0.161
0.323
0.123
0.331
t26-t28
-0.238
0.553
0.237
0.575
-0.256
0.551
0.045
0.559
constan
t-1.495
0.234
***
-0.877
0.663
-1.344
0.240
***
-1.173
0.402
***
Tab
le10:Baselinehazard,ap
plication
toR&D
42
Variab
leCoeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
t2-t4
-0.568
0.168
***
-16.857
0.323
***
-0.560
0.168
***
-0.492
0.171
***
t4-t6
-0.467
0.172
***
0.068
0.552
-0.457
0.172
***
-0.268
0.182
t6-t8
0.095
0.144
-0.322
0.425
0.115
0.144
-0.057
0.152
t8-t10
0.137
0.148
-0.205
0.601
0.154
0.148
-0.281
0.161
*t10-t12
0.247
0.139
*-0.728
0.712
0.271
0.139
*-0.154
0.157
t12-t14
-0.162
0.160
0.327
0.590
-0.139
0.160
-0.453
0.184
**
t14-t16
-0.680
0.214
***
0.095
0.585
-0.656
0.214
***
-0.847
0.237
***
t16-t18
0.108
0.161
-0.850
0.877
0.133
0.161
-0.303
0.183
*t18-t20
0.126
0.195
-0.204
0.620
0.165
0.196
-0.263
0.206
t20-t22
-0.666
0.293
**-0.216
0.765
-0.623
0.294
**-0.956
0.314
***
t22-t24
-1.056
0.413
**-16.361
0.764
***
-1.025
0.412
**-0.944
0.438
**
t24-t26
-1.509
0.558
***
-16.565
0.493
***
-1.484
0.556
***
-0.869
0.585
t26-t28
-1.832
1.039
*-17.194
0.878
***
-1.815
1.041
*-1.220
1.038
postdocxt2
-t4
-11.919
1.027
***
17.502
0.515
***
-11.949
1.027
***
-11.913
1.026
***
postdocxt4
-t6
3.207
1.271
**3.213
1.268
**3.302
1.278
***
postdocxt6
-t8
-12.688
1.117
***
0.340
0.804
-12.752
1.111
***
-12.954
1.136
***
postdocxt8
-t10
-12.917
1.086
***
0.794
0.748
-12.980
1.087
***
-13.256
1.094
***
postdocxt10-t12
-12.943
1.129
***
0.642
1.258
-12.953
1.124
***
-13.131
1.114
***
postdocxt12-t14
-12.300
1.112
***
-12.333
1.109
***
-12.404
1.105
***
postdocxt14-t16
-12.499
1.236
***
-0.819
1.387
-12.535
1.224
***
-12.702
1.274
***
postdocxt16-
t18
1.639
1.109
1.687
1.122
1.134
1.115
postdocxt18-t20
3.370
1.050
***
3.203
1.050
***
2.687
1.044
***
postdocxt20-t22
-11.355
1.284
***
1.976
1.198
*-11.561
1.283
***
-11.886
1.310
***
postdocxt22-t24
-13.814
1.508
***
-13.982
1.507
***
-14.752
1.518
***
constant
-4.434
0.532
***
-18.843
0.400
***
-4.721
0.543
***
-7.089
1.182
***
Tab
le11:Baselinehazard,R&D
tonon
-S&E
43
Variable
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
Coeff
St.Err
t2-t4
-0.023
0.110
-0.087
0.375
-0.023
0.110
-0.019
0.110
t4-t6
-0.138
0.126
-0.725
0.499
-0.138
0.126
-0.038
0.125
t6-t8
-0.124
0.133
-1.923
1.041
*-0.126
0.133
-0.176
0.134
t8-t10
0.018
0.129
-0.020
0.465
0.013
0.129
-0.169
0.132
t10-t12
-0.102
0.151
-0.776
0.709
-0.108
0.152
-0.191
0.154
t12-t14
-0.353
0.179
**
-14.515
0.517
***
-0.358
0.179
**
-0.379
0.182
**
t14-t16
-0.288
0.198
0.482
0.779
-0.299
0.197
-0.320
0.209
t16-t18
0.390
0.171
**
-0.589
1.105
0.383
0.172
**
0.165
0.184
t18-t20
0.614
0.230
***
-14.544
0.511
***
0.604
0.232
***
0.284
0.238
t20-t22
-0.564
0.465
-14.541
0.579
***
-0.569
0.468
-0.641
0.452
t22-t24
-1.498
1.004
-15.176
0.917
***
-1.511
1.006
-1.279
0.979
t24-t26
0.789
0.315
**
0.534
0.573
0.777
0.313
**
1.333
0.344
***
t26-t28
-10.812
0.322
***
-13.619
1.030
***
-12.427
0.319
***
-11.504
0.358
***
constant
-4.588
0.572
***
-18.423
4.021
***
-4.720
0.586
***
-5.463
1.030
***
Tab
le12:Baselinehazard,ap
plication
tonon
-S&E.
44
Appendix
Occupational Categories in Sciences and EngineeringMajor category Minor categories of occupation
Sciences and EngineeringComputer Computer & Information Scientists& Mathematical Mathematical ScientistsScientists Postsecondary Teachers - Computer & Mathematical SciencesLife Scientists Agricultural & Food Scientists
Biological ScientistsEnvironmental Life ScientistsPostsecondary Teachers - Life & Health Sciences
Physical Scientists Chemists, except BiochemistsEarth Scientists, Geologists & OceanographersPhysicists & AstronomersOther Physical ScientistsPostsecondary Teachers - Physical Sciences
Engineers Aerospace & Related EngineersChemical EngineersCivil & Architectural EngineersElectrical & Related EngineersIndustrial EngineersMechanical EngineersPostsecondary Teachers - EngineeringNon-Sciences and Engineering
Non-S&E occupations Managers & AdministratorsNon-S&E Postsecondary TeachersSocial Services & Related occupationsTechnologists & TechniciansSales & Marketing occupationsArt, Humanities & related occupations
SOURCE: SESTAT: A Tool for Studying Scientists and Engineers in the United States.Division of Science Resources Studies. NSF.
45