67 T.C. Buchmueller et al. Economics of Education Review 14 1999 65–77
the student collaborated with this individual, and a meas- ure of the “selectivity” of the school from which the
baccalaureate was obtained. Each graduate education variable has the expected relationship with the quality of
the first post-Ph.D. job as well as with subsequent research productivity. Although prior-to-Ph.D. research
output shows a strong, positive effect on subsequent out- put, it is only weakly related to job placement. Long’s
results are essentially corroborated by Long and McGinnis 1981.
3. Empirical specification
The preceding discussion suggests the following model of the determination of publishing proficiency:
PUBS 5 ƒ
TRAINING, JOB, X, u 1
where an economist’s post-Ph.D. publication record PUBS depends on his graduate training TRAINING,
the type of job he holds immediately after completing that training JOB, his observed personal characteristics
X and unobserved factors u. Our data on publications, training, employment, and personal characteristics are
discussed in Section 4.
One problem in estimating this model is that the indi- vidual’s first job is endogenous. That is, new Ph.D. econ-
omists differ in their research interests and publishing potential. Future productivity is obviously uncertain at
the time they enter the job market, but prospective employers have more information on which to base a
prediction than do we as analysts. Thus, we would expect that individuals with higher than average publishing
potential will sort into jobs that reward publishing e.g. research-oriented academic positions. Failing to account
for this endogeneity means that our estimates of the relationship between job placement and early publi-
cations will overstate the true structural impact of the former on the latter.
With data on factors that influence job placement but not directly influence subsequent research output, we
could estimate a linear version of Eq. 1 by instrumental variables. Unfortunately, however, we lack such ident-
ifying variables.
3
As an alternative to structural esti-
3
In earlier versions of this study we attempted to use the year an individual went on the job market to identify the model.
The assumption was that year dummies would pick up the effect of demand shifts which influenced the number and type of jobs
available, but were uncorrelated with any individuals’ charac- teristics. The problem is that our sample is defined by the year
an individual received his or her Ph.D., so the year an individual went on the job market is correlated with the length of time in
graduate school and the distance from completing the disser- tation. A GMM orthogonality test Newey, 1985 rejected the
validity of the year dummies as instruments.
mation, we take the following approach. First, we esti- mate best linear predictors of publications given
individual and graduate program characteristics, but not employment characteristics. These results may be inter-
preted as reduced form parameter estimates from a sys- tem of structural equations determining both job place-
ment and publishing proficiency. As such, these estimates capture the covariation of publications and pre-
dictor variables associated with both the direct and the indirect effects of these variables on realized publi-
cations output.
We also estimate best linear predictors in which a measure of job placement is included among the predic-
tor variables. For reasons discussed below, we divide jobs held by Ph.D. economists into two categories: fac-
ulty jobs at research universities AcademicPh.D. sector and all other jobs. The estimated relationship
between job type and publications combines the struc- tural impact of an individual’s employment setting on
publications and the effect of unobserved heterogeneity among factors influencing selection into employment
sector. Finally, to explore this relationship further and to examine how early research experience influences job
placement, we estimate a probit model in which the dependent variable equals one for individuals whose first
jobs are in the AcademicPh.D. sector and equals zero for individuals who sort into other types of positions.
An additional consideration in the estimation of Eq. 1 is that the empirical distribution of publications is
discrete, with large probability mass at zero. One poss- ible econometric specification to accommodate count
data is the compound Poissonnegative binomial model Cameron and Trivedi, 1986; Goodwin and Sauer, 1995.
Other authors have estimated a Tobit model of publi- cations Levin and Stephan, 1991.
In preliminary work we estimated publications regressions by Tobit, negative binomial, and least
squares. We also estimated probit models in which the dependent variable equals one for individuals with one
or more publications. The qualitative results — i.e. the sign and significance of key parameter estimates — did
not vary substantially across the different specifications. We prefer restricting attention to the least squares esti-
mates for several reasons. The Tobit and Poisson models rely on strong distributional assumptions that are parti-
cularly problematic in the presence of the endogeneity bias discussed above. Moreover, the Tobit model is only
appropriate if we were to treat the publications variable as the observed counterpart of a censored latent variable,
such as individual productivity. In contrast, interpretation of least squares estimates as best linear predictors of the
number of publications requires no assumptions on the structure of the process generating publications. Before
turning to the regression analysis, we describe our data and present descriptive results concerning the relation-
68 T.C. Buchmueller et al. Economics of Education Review 14 1999 65–77
ships among graduate training, job placement, and publi- cation productivity.
4. Data