Fishes, ponds, and productivity: student-advisor matching and early career publishing success for economics PHDS.

Hilmer, Michael J. ; Hilmer, Christiana E.

I. INTRODUCTION

Research productivity is an important concern for members of the academe. Publishing success is important for individual departments because more productive faculty increases the department's research profile, which in turn increases the department's professional reputation. As evidence, Thursby (2000) and Ehrenberg and Hurst (1998) find that a department's reputation in the National Research Council (NRC) ratings increases with published research, while Smyth (1999) finds that pages published in the top 5 journals have a greater impact than pages published in other journals. Publishing success is important for individual faculty members because peer-reviewed publications have long been shown to be economic currency for academic economists (Bratsberg, Ragan, and Warren 2003; Moore, Newman, and Turnbull 1998; Sauer 1988 among others). Research promise is important for graduate students because students with the greatest promise for productive careers are likely to receive the best initial job placements (Buchmueller, Dominitz, and Hansen 1999; Krueger and Wu 2000; Long 1978). Indeed, according to a survey of hiring departments (Carson and Navarro 1988), every single top 20 domestic economics program considered the quality of one's research agenda to be of moderate to great importance in the hiring decision.

An important question then is how to identify which factors are associated with the likelihood of having a more productive research career. To date, empirical studies have focused on the program from which the student graduated (recent examples include Collins, Cox, and Stango 2000; Davis, Huston, and Patterson 2001; Stock and Alston 2000). This focus implicitly assumes that because only the very best students become enrolled in elite programs, graduates of highly ranked programs should hold the greatest promise for publishing success. (1) In support, Coupe (2001) and Buchmueller, Dominitz, and Hansen (1999) find that students graduating from top programs are more likely to publish in core economics journals, while Hogan (1981) finds that students graduating from programs with more active faculty publish more journal articles.

Before placing total faith in these results, however, members of the academe should ask whether such findings hold across all points in the graduate distribution or whether they are only accurate predictors on average. Previously published data clearly suggest the latter. For instance, Buchmueller, Dominitz, and Hansen (1999) find that 30% of Tier 1 graduates do not publish any academic articles within the first 9 yr after PhD receipt, while 50% do not publish any top 50 articles (as defined by Liebowitz and Palmer 1984). In other words, there is a "large amount of unexplained variability" in the future success of a program's graduates (Krueger and Wu 2000, 93). Consequently, it is likely that significant overlap exists between the performance of top program graduates and lesser program graduates, with significant portions of the latter outperforming significant portions of the former.

With this in mind, the question becomes: is there a readily observable factor that is a less noisy predictor of a student's likelihood for publishing success? We argue that there is. During a student's postgraduate education, one of the most significant influences is his or her dissertation advisor. We posit that for timing reasons, advisors will possess additional insight into a student's true research potential that was not available to the admission committee when deciding whether to admit the student into the program. Consequently, we expect a student's dissertation advisor to be a more informed predictor of his or her early career productivity than the program from which he or she receives the PhD. (2) In other words we believe that considering the identity of a student's dissertation advisor in addition to the quality of the program from which the student receives the PhD should help reduce the uncertainty inherent in predicting the student's position within the early career productivity distribution.

The current study is the first to address this possibility. Using readily available data sources, we are able to uniquely identify a student's primary dissertation advisor as well as his or her graduate program, dissertation field, sex, domestic/international status, first postgraduate job, and several different measures of his or her early career research productivity. Based on Coupe's (2003) global rankings of the top 1,000 economists, we are able to assign the advisor's "ranking" within the profession. Combining these factors, we are able to examine the relationship between the ranking of a student's dissertation advisor and the student's early career research productivity.

We analyze a sample of 1,892 individuals receiving economics PhDs from top 30 programs during a 5-yr period in the early 1990s. Our main finding is that, controlling for program quality, student-advisor match is a significant predictor of early career research productivity, especially for publications in top 36 economics journals. Additionally, controlling for advisor rank significantly reduces the estimated productivity differences due to program quality, ceteris paribus, suggesting that much of the estimated productivity difference previously attributed to differences in program quality might actually be explained by differences in the student-advisor match. (3) Finally, simulations based on our regression results suggest that potentially significant overlaps exist in the cross-program productivity distribution, as we find that while Tier 1 graduates with star advisors are statistically more productive than everyone else, Tier 2 and Tier 3 graduates with star advisors perform as well as Tier 1 graduates with lower ranked advisors and Tier 2 and Tier 3 graduates with lower ranked advisors perform as well as Tier 1 graduates with unranked advisors.

[FIGURE 1 OMITTED]

II. THEORETICAL BACKGROUND

Our central prediction is that students with more prominent dissertation advisors will be more productive researchers in their early careers than otherwise similar students with less prominent advisors. We make this argument based on observations about the timing of the process by which economics PhD students become matched with their dissertation advisors as opposed to their PhD programs. We start by noting that thousands of potential students apply to PhD programs in economics each year. (4)Presumably, each program desires to enroll the best possible class of entering PhD students. While several different factors might go into the determination of the best class (for a detailed discussion, see Attiyeh and Attiyeh 1997), Krueger and Wu (2000, 81) state that "one important consideration is the eventual job placement and professional success of their graduates." In a world with perfect information about a student's research ability, students with the highest potential for future success should be admitted to the best programs. The information available when making admission decisions is imperfect, however, as it is mostly based on the student's standardized test scores, prior academic performance, and letters of recommendation. As Cushing and McGarvey (2004) argue, such observable measures of a student's potential likely vary little across advanced-degree aspirants, as those desiring to further their educations are mostly drawn from the top tail of the student distribution. In other words, because nearly all applicants, especially to top programs, are high achieving students with top references, there will be "considerable uncertainty in forecasting which applicants will be successful economists" (Krueger and Wu 2000, 93).

This uncertainty likely results in "errors" in the admissions process, whereby many students who fail to publish in their early career have their PhDs minted from top programs, while many students who are relatively prolific in their early careers have their PhDs minted by lesser programs. This overlapping distribution of early career research productivity is clearly demonstrated in Figure 1 for the data set we analyze below. Given that we analyze a sample of students receiving PhDs from top 30 economics programs between 1990 and 1994, the box plots should be interpreted as follows: the ends of the boxes represent the number of articles published per year between PhD receipt and December 2002 by students at the 25th and 75th percentiles within Tier 1, Tier 2, and Tier 3 programs, respectively, while the line in the middle of the box represents the number of articles published by the median student and the right-most whisker represents the number of articles published by the student at the top end of the distribution below outside values. As such, the box plots indicate that there is significant overlap in the distributions of early career research productivity across graduates from different program tiers. In other words, it should be clear from Figure 1 that program rank alone is not a sufficiently accurate predictor of a student's likelihood for early career publishing success.

Consider now the process through which PhD students become matched with their dissertation advisors. A student normally seeks out an advisor only after he or she has completed significant amounts of coursework, taken and passed preliminary exams, and potentially worked for several terms as a research assistant. As such, by the time a student formally requests that a given faculty member serve as his or her advisor, the faculty member will have access to significant additional information as to the student's likely research potential. Because of this additional readily observable information, we expect advisors to be much better informed as to a given student's likelihood for future publishing success than the admission committee was when deciding to admit the student into the program. As a result, the signal provided by the student-advisor match should be less noisy than the signal provided by the student-program match, and we would expect students with more prominent advisors to be more productive in their early careers than otherwise similar students with less prominent advisors.

As with most educational outcomes, there are two potential reasons why students matched with more prominent advisors might become more productive researchers. The first is a human capital argument (Becker 1964). Namely, more prominent economists are more prominent because they themselves are prodigious publishers. It is therefore possible that they pass along some of their productive knowledge to their advisees. The other is a signaling argument (Spence 1973) that by becoming paired with a prominent advisor, graduate students are revealing themselves to possess the characteristics that will make them successful publishers in their early careers. We note that this may result from either advisors screening out lower ability/less motivated students or lower ability/less motivated students choosing to work with less demanding faculty members. For the purposes of this study, it is not necessary to endorse one of these explanations over the other, as they both predict greater success for students working with highly ranked advisors. Nonetheless, a very interesting question for future research might be whether our predicted outcome is due to a signaling or human capital effect.

III. DATA

A major innovation of this study is the construction of a first-of-its-kind data set that matches economics PhD recipients to their dissertation advisors, peer-reviewed publication records, graduate programs, dissertation fields, sex, domestic/international status, and first postgraduation jobs. The Dissertation Abstracts database (published by ProQuest Information and Learning) contains extensive information on more than 1.2 million dissertations accepted at accredited North American educational institutions since 1861. In 1990, the database started including the name of the student's dissertation advisor for the vast majority of dissertations filed. (5) We collected information on 1,892 dissertations filed in economics fields between 1990 and 1994 for students graduating from top 30 economics programs and reporting the identity of their dissertation advisor. To make sure that we do not include students writing on economic topics but belonging to different academic disciplines, we cross-reference our list with the "Doctoral Dissertations in Economics Annual List" published each December in the Journal of Economic Literature. The 5-yr time frame is somewhat arbitrary but is chosen for the following reasons. We begin in 1990 because that is the first year in which dissertation advisors are included in the majority of student records. We collect data over multiple years to avoid any single-year aberrations that might bias the results. (6) Finally, we cut off the time frame in 1994 because tenure is usually granted after a faculty member's sixth year, and therefore, at the time we collected the data, we were observing the student's productivity up to the point where their initial tenure decision was made (with a lag to allow for time in press).7

Individual-specific peer-reviewed publication data as of December 2002 are collected from Econlit, which is the American Economic Association's bibliography of economics literature throughout the world. The database contains information on articles published in more than 700 journals, including all the major field and general interest economics journals. In other words, while some publications may not be contained in Econlit, they are likely published in more obscure, less respected journals, or in the words of Coupe (2003, 1310), "one can claim with a slight exaggeration, first, that if one is not in Econlit, one did not do academic research in economics and second, that these journals together form the 'economics literature'." To define research productivity, we consider three commonly used metrics. The first two are the total number of articles published and the total number of articles published in top 36 economics journals according to Scott and Mitias (1996). The third is included to address the concern that "an article is not an article" and follows Sauer (1988) by calculating a measure of pages published that is weighted for journal quality, number of authors, and number of characters per page (AEQ pages).

To rank economics programs, we follow the three-tier ranking presented in Siegfried and Stock (2001). The three tiers correspond to programs 1-6, 7 15, and 16 30, respectively, in the 1995 NRC rankings of PhD-granting economics programs. (8)

To rank dissertation advisors, we use the global top 1,000 economist ranking of Coupe (2003). This ranking is based on a weighted average of 11 different historically used metrics of research productivity. (9) By calculating a weighted average of these metrics, each of which was developed in response to perceived weaknesses in previous methodologies, Coupe is hoping to avoid the complaint that "we were disadvantaged by the specific weighting scheme." Overall, we define an advisor as being ranked among the worldwide top 250 ("star" advisors), ranked between 251 and 1,000 ("lower ranked" advisors), or not ranked in the top 1,000 ("unranked" advisors).

Finally, as previous studies by Davis, Huston, and Patterson (2001), Collins, Cox, and Stango (2000), and Buchmueller, Dominitz, and Hansen (1999) indicate, an important determinant of a student's future productivity is whether he or she holds a research-oriented job. We define research-oriented jobs as those in the academic sector or with the Federal Reserve. (10) To determine a student's first postgraduation job, our initial source is the self-reported information contained in the American Economic Association's Directory of Members. For students whose information was not listed, we turn to the author affiliation in Econlit for the first article published after the student received his or her PhD.

IV. DESCRIPTIVE ANALYSIS

A. Student-Program and Student-Advisor Match

Table 1 provides a summary analysis of several different aspects of the student-program and student-advisor matchings. Overall, roughly 28%, 39%, and 33% of the students in our sample received their PhDs from Tier 1, Tier 2, and Tier 3 programs, respectively. Given the different number of programs in each tier, we observe more graduates of top programs, with an average of roughly 17.5 students per year observed graduating from each Tier 1 program as opposed to averages of roughly 16.6 and 8.4 observed graduating from each Tier 2 and Tier 3 programs. The higher concentration of students within top programs is not surprising given Coupe's (2001) estimate that 80% of PhDs in economics graduate from just 20 programs and the finding of Pieper and Willis (1999) that top 10 schools produce 47% of the economics faculty at PhD-granting institutions. As might be expected, there appears to be an increased supply of ranked advisors within Tier 1 programs, as nearly 65% of Tier 1, 47% of Tier 2, and 32% of Tier 3 advisors are ranked among Coupe's top 1,000 worldwide economists.

Turning to the middle panel, the clear plurality of students, roughly 41%, work with unranked dissertation advisors. This is not surprising given that there are far more than 1,000 global academic economists. (11) Nonetheless, nearly three-fifths of our students did have their dissertations directed by a top 1,000 advisor. Turning to the distribution of advisors, among those we observe directing at least one dissertation, roughly 18% are stars, 29% are lower ranked, and 53% are unranked.

The bottom panel of Table 1 combines student demand and advisor supply by presenting the cross-distribution of students by program tiers and advisor ranking groups. While we find that the percentage of students working with star faculty decreases with program tier (50%, 22%, and 18%), our summary statistics suggest that there are significant cross-tier overlaps in the quality of advisors with which students are able to work. Specifically, while 18% of Tier 3 students work with star advisors, 50% of Tier 1 and 78% of Tier 2 students work with either lower ranked or unranked advisors.

Table 2 presents the distribution of completed dissertations directed across all advisors we observe lead supervising at least one completed dissertation between 1990 and 1994. Overall, we observe 741 different lead advisors with 341, or nearly 46%, being ranked among Coupe's top 1,000. (12) A clear majority of advisors maintain lighter loads, with nearly 47% lead supervising only one dissertation and nearly 69% averaging one or fewer per year. At the opposite end, six faculty members supervise an average of three or more dissertations per year with the maximum number of advisees being 21, or 4.2 students per year. It further appears that ranked advisors tend to carry larger advising loads, as the percentage of advisors who are ranked in the top 1,000 increases with the number of total advisees supervised. These trends are broadly consistent with those observed in a study of all dissertations filed at Cornell University from 1996 to 2002 (Crosta and Packman 2005). During that period, social science faculty chaired an average of 1.13 dissertations with the top 10% chairing 55% of all dissertations. Van Ours and Ridder (2003) explain at least a part of this trend by observing that in the Netherlands, dissertation supervisors with good research records are paired with better students and are thus more likely to supervise completed dissertations than faculty with lesser records who are paired with students who are more likely to stop short of completion.

B. Early Career Productivity

To move the analysis toward early career productivity, Table 3 presents average values across program tier and advisor rank for each of our productivity metrics. Overall, by December 2002, our sample of 1990-1994 PhD recipients averaged 4.12 total articles, 0.35 top 5 articles, 1.30 top 36 articles, and 13.69 AEQ pages. These publication data appear generally consistent with previously published data (Buchmueller, Dominitz, and Hansen 1999; Coupe 2001), as roughly 68% of our students publish at least one article within their first 8 12 yr after graduation. They also illustrate the difficulty associated with publishing in the very best general interest journals, as only slightly less than 40% of our students are able to publish a top 36 article in their early careers (while less than one in seven are able to publish in a top 5 journal).

The middle panel of Table 3 suggests that large cross-tier differences exist, with Tier 1 graduates being, on average, more productive across all metrics than Tier 2 and Tier 3 graduates. Perhaps most significantly, these average differences are largest for the highest quality publications. Namely, while Tier 1 students average roughly 50% more total articles than Tier 2 and Tier 3 students, they average more than twice as many AEQ pages and more than 2.2 times as many top 36 articles. In other words, it appears that the biggest difference between students graduating from elite programs and students graduating from lesser programs occurs in the propensity to publish in the very best economics journals. Finally, the bottom panel suggests similar patterns for students with star advisors relative to those with lower ranked and unranked advisors.

V. EMPIRICAL RESULTS

The next step in our analysis is to empirically assess the degree to which the rank of a student's dissertation advisor affects his or her early career productivity. To isolate this effect, we estimate productivity functions for each of our three metrics that control for the rank of a student's dissertation advisor, the rank of his or her PhD program, and other individual characteristics. Following standard form, our estimation equations can be written as:

(1) [P.sub.i] = [B.sub.0] + [B.sub.1][A.sub.i] + [B.sub.2][Q.sub.i] + [B.sub.3][X.sub.i] + [B.sub.4][O.sub.i] + [[epsilon].sub.i],

where [P.sub.i] represents one of the four productivity measures, [A.sub.i] is the rank of the student's dissertation advisor, [Q.sub.i] is the reputation rank of the student's PhD program, [X.sub.i] is a vector of individual characteristics, and [[epsilon].sub.i] is an error term. The individual characteristics we consider are whether the student is male or international, the field in which the student's dissertation is filed, the number of years since the student received his or her PhD, and whether the student's first job was research oriented. As demonstrated in Table 2, advisors differ greatly in their propensity to take on advisees. The number of other advisees with whom a student's advisor works might have competing effects on his or her future productivity because, on the one hand, the increased student load could force the advisor to devote less time to each student, thereby harming the student's learning. On the other hand, anecdotal evidence suggests that prominent advisors might take on increased student loads due to their love of mentoring, and thus, they may actually devote more time to each of their students than would have other advisors with smaller student loads. To account for these possibilities, our vector of individual characteristics also includes [O.sub.i], which indicates the number of other completed dissertations that the student's advisor directed during our sample period. Our main parameters of interest are [B.sub.1] and [B.sup.2], which indicate the effect that the rank of a student's dissertation advisor and the reputation rank of a student's PhD program have on his or her early career productivity, all else constant.

We note two important estimation concerns associated with our empirical approach. First, our total and top 36 article measures are count data and are truncated at 0 due to the fact that many students have not published, especially within top 36 journals. Truncated count data models are normally estimated as either a Poisson or a negative binomial, both of which account for the skewed distributions of the dependent variables (Cameron and Trivedi 1998). A well-known problem with the Poisson distribution is the presumed equality of the conditional mean and variance functions. The data in our analysis fail tests of overdispersion for each productivity measure, suggesting that the assumption of equidispersion is violated and that the Poisson is not the appropriate distribution. As a result, we estimate each of our count data productivity functions with the negative binomial regression model as that distribution accounts for the skewness of the data without requiring equality between the conditional mean and variance.

Second, as noted by Buchmueller, Dominitz, and Hansen (1999) and others, a student's initial job placement is endogenous because the initial placement that a student receives likely influences his or her need and/or desire to publish in the early career. Failing to account for this endogeneity would cause us to overestimate the true relationship between the student's initial job placement and his or her early career productivity. The ideal correction for such potential endogeneity would be an instrumental variables approach. Unfortunately, as with Buchmueller, Dominitz, and Hansen (1999), we lack readily observable factors that would influence the student's initial job placement without also potentially influencing his or her early career productivity. As a result, we are unable to employ the desired instrumental variable correction. Nonetheless, we do not want to totally ignore the potential influence of the endogenous initial job placement. We, therefore, employ an approach similar to that in Buchmueller, Dominitz, and Hansen (1999) whereby we estimate our productivity functions both with and without the first-job research-job variable in an effort to gauge the impact that the relationship between a student's advisor and his or her initial job placement has on the student's early career publishing success. (13)

A. Are Students with Higher Ranked Advisors More Productive?

To examine the effect that the student-advisor match has beyond the initial student-program match, Table 4 presents results of estimating Equation (1) for each of our three productivity measures that have been converted to marginal effects. The first three columns present results that do not control for initial job placement, while the last three present results that include the first-job research-job variable. In every column, advisor rank and program rank are entered as sets of dummy variables, with the omitted group being Tier 3 graduates with unranked advisors. Hence, the marginal effects presented in Table 4 represent the estimated differences in each of our productivity measures for students having an advisor belonging to a given ranking group or graduating from a given program tier relative to otherwise similar Tier 3 students with unranked advisors.

Comparing the first three columns of results to the second three suggests that at least part of the impact that advisor rank (and program tier) has on early career publishing success takes place through the impact that advisor rank (and program tier) has on the likelihood of receiving an initial placement in a research-oriented job. Specifically, after including the first-job research-oriented variable, the estimated effect of advisor rank decreases by up to one-third, while the estimated effect of program tier decreases by more than one-half. Consequently, the discussion below focuses on the results that control for the type of initial placement that a student receives.

While not presented here for the sake of brevity, comparing these results to results that include program tier but not advisor rank suggests two major findings. First, after adding controls for the global rank of a student's advisor, the estimated differences between Tier 1, Tier 2, and Tier 3 graduates decrease in magnitude by roughly one-third for each metric. At the same time, the estimated log likelihoods increase by amounts large enough to suggest that our controls for advisor rank are statistically significant. (14) These results are consistent with our central hypothesis, as they suggest that significant portions of the difference between graduates of top- and lower ranked programs might be explained by the match between the student and his or her dissertation advisor.

Second, and most importantly for this analysis, after controlling for the quality of program from which a student graduates, students with ranked advisors, and especially those with star advisors, are statistically more likely to publish across all metrics than students with unranked advisors. Specifically, holding program quality constant, we estimate that students with star advisors produce 1.45 more total articles, 0.87 more top 36 articles, and 9.69 more AEQ pages than otherwise similar students with unranked advisors, while students with lower ranked advisors produce 0.65 more total articles, 0.62 more top 36 articles, and 6.74 more AEQ pages than otherwise similar students with unranked advisors. Hence, the estimated differences suggest that the student-advisor match provides a strong signal as to whether the student will publish any articles and an especially strong signal as to the likelihood that a student will publish in top economics journals early in his or her career.

Turning to the remaining variables, our results suggest that, controlling for program reputation, years since PhD receipt, and domestic/international status, males are statistically likely to publish more total and top 36 articles but not AEQ pages than otherwise similar females, while, all else constant, international students are less likely to publish across all three metrics than otherwise similar domestic students. These results are consistent with previous findings in Buchmueller, Dominitz, and Hansen (1999). The number of other advisees that a student's advisor supervises during our 5-yr period is never estimated to have a statistically significant effect on the student's early career productivity. Finally, while not presented here, as with Buchmueller, Dominitz, and Hansen (1999), we find few statistically significant differences across fields of study, suggesting that, all else equal, a student's chosen field does not have much impact on his or her early career productivity.

B. Fishes and Ponds: Do Cross-Program Productivity Distributions Overlap?

As a final exercise, we calculate predicted productivity measures for a hypothetical student under each of the possible advisor rank/ program tier combinations. In this case, our hypothetical student is a male, domestic student holding a research job, having had his PhD for the sample average 9.81 yr, and having an advisor who supervised the sample average 3.91 other completed dissertations. In essence, the results in Table 5 replicate the experiment of: (a) sending the hypothetical student to a program in each quality tier and having him work with a star, a lower ranked, and an unranked advisor within each of those programs, (b) having the student attain a research-oriented job, (c) observing his early career productivity in each instance, and (d) comparing the results.

A notable finding emerges from Table 5. Figure 1 suggests that there is significant overlap in the productivity distributions of students graduating from Tier I, Tier 2, and Tier 3 programs. The predicted productivity measures in Table 5 suggest that the student-advisor match might help reduce the noise associated with determining a student's research potential. Namely, outcomes for our hypothetical student fall into four statistically different groups. The student is predicted to be most productive across all four metrics if he attends a Tier 1 program and works with a star advisor. This is not at all surprising. What is potentially surprising is the pairings belonging to the remaining outcome groups. The second most productive grouping is for star advisors at Tier 2 programs, lower ranked advisors at Tier 1 programs, and star advisors at Tier 3 programs. The third most productive grouping is for unranked advisors at Tier 1 programs, lower ranked advisors at Tier 2 programs, and lower ranked advisors at Tier 3 programs. The least productive grouping is for unranked advisors at Tier 2 and Tier 3 programs. In other words, Tier 2 and Tier 3 graduates working with star advisors are predicted to do as well as all Tier 1 graduates not working with star advisors, while Tier 2 and Tier 3 graduates working with lower ranked advisors are predicted to do as well as Tier 1 graduates working with unranked advisors. Put another way, it appears that there are potentially tangible benefits to being a "big fish in a small pond" in that students might perform better if they attend a lower ranked program but are able to work with a more prominent advisor.

C. Are the Results Robust to Different Specifications of Advisor Rank and Program Tier?

Given that advisor quality and program rank could clearly be entered into our productivity functions in a number of different ways (for instance, as linear measures, series of dummy variables, or a number of quadratic terms), we believe that it is important to note why we choose our specific three-tier definition of advisor rank and program tier. Entering advisor rank and program tier as linear measures requires the assumption that there are distinct, constant differences between advisors (and programs) of a given rank and advisors (and programs) who are ranked either one position higher or one position lower. Given the imperfect science associated with quantifying research productivity (as evidenced by the extensive literature attempting such rankings), we believe that it is possible to quibble over whether a given individual (or program) should be ranked, say 17th or 18th. At the same time, however, we believe that broader groupings are highly accurate in terms of relative research productivity. This feeling is similar to that of Kingston and Smart (1990, 149) who suggest that such a categorical approach is preferable to a linear specification when comparing graduates of different-quality colleges because "it is likely that differences throughout most of the academic hierarchy are inconsequential [which would imply only a small overall effect of program rank] ... but that going to an elite school does make a difference." This is the very argument that Stock and Alston (2000) employ to motivate their use of the three-tier program quality measure in their analysis of the effect of program quality on initial job placements and is presumably why other studies, such as Buchmueller, Dominitz, and Hansen (1999), employ a categorical approach to defining program quality.

While not presented here, we note that our results appear to be robust across the numerous different specifications of the advisor rank and program tier that we estimated. For example, in a regression run only for students with star or ranked advisors, when entered as a dummy variable, our estimated marginal effect for star advisors is 0.8911. When entered as a continuous measure, on the other hand, the estimated marginal effect is 0.00162, suggesting that, all else equal, every one position increase in an advisor's relative standing is associated with his or her student averaging 0.00162 more total articles. Given that the middle ranking in the star advisor category (advisors ranked 1-250) is 500 positions above the middle ranking in the ranked advisor category (advisors ranked 251-1,000), the results suggest that a student of the former-type advisor would average 0.8137 more total articles than a student of the latter-type advisor, ceteris paribus.

Another potential concern is that our categorical approach to defining program rank might be masking potential cross-program differences in the relationship between the advisor-student match and the early career publishing success. For example, because Harvard has significantly more advisors ranked in Coupe's top 1,000 (49 as opposed to 34 for Chicago, 29 for MIT, 27 for Stanford, 23 for Princeton, and 18 for Yale), it could be possible that the statistical significance of the advisor rank variable for Tier 1 programs is simply picking up some sort of "Harvard effect." To address this concern, we estimated separate productivity functions that included an exhaustive set of 29 program dummies (Harvard omitted) rather than our categorical variables. Again, the results were nearly identical in significance and magnitude to those reported in the text. Moreover, simple correlations between advisor rank and student productivity for every program with sufficiently large numbers of student observations yield the expected positive correlations.

Finally, to close our analysis, we perform two additional robustness checks. First, to investigate the possibility that the estimated relationships might differ across program tiers, we estimated the models separately for each program tier, finding that the star and lower ranked advisor variables were positive and statistically significant within each program tier for all three productivity measures. Second, to loosen the restriction that the effect of advisor rank is the same across the different program tiers, we estimated productivity functions that included interaction terms between advisor rank and program tier, finding that the interaction terms were statistically insignificant and that their inclusion did not affect the estimated relationship between advisor rank and early career productivity in all cases.

VI. CONCLUSIONS

This paper is the first to examine the effect that the student-advisor match has on a student's early career productivity. Regression results confirm the significance of working with a ranked dissertation advisor, as we find that students working with ranked advisors average significantly more publications, especially in terms of top 36 articles and AEQ pages, than students working with unranked advisors, ceteris paribus. This result holds even after controlling for the first job that a student holds. Consequently, it appears that the "quality" of a student's dissertation advisor is an important predictor of early career success beyond the reputation of the program from which the student graduates. Moreover, in many cases, this advisor effect appears to outweigh the school effect to such an extent that students attending lower ranked programs but working with superstar advisors are predicted to publish significantly more total and top 36 articles and more quality-adjusted pages than students attending top-ranked programs but working with less prolific advisors.

Our results are potentially important for economics departments that are considering which applicants to pursue on the job market, especially those lower ranked departments that might be choosing between lower distribution students from top programs and higher distribution students from lower ranked programs. As a specific example, Smyth (1999) estimates that by publishing one additional 10-page article in a top 5 journal, the Department of Economics at Louisiana State University could increase its NRC ranking from 51 to 40 and its overall professional perception from "adequate" to "strong." Our results suggest that by hiring a Tier 2 student with a star advisor instead of a Tier 1 student with an unranked advisor, the department could increase its productivity by nearly 5 AEQ pages, thereby significantly increasing its professional reputation.

Not only are our results important to hiring committees, but they are potentially important to current and potential economics PhD students as well. Namely, our results suggest that students attending lower ranked programs may outperform students attending top 6 programs if they are able to work with a superstar advisor. Given the importance that hiring departments place on research potential (Carson and Navarro 1988), these results suggest that students might actually be better off attending lower ranked programs and having the opportunity to work with more respected advisors than by attending a top program, falling in the lower tails of the student ability distribution, and being left to work with a lesser known advisor. In other words, it appears that many students might benefit by being a "big fish in a small pond" rather than a "small fish in a big pond." These results are also likely to translate into labor market success. For instance, Sauer (1988) estimates that each additional AEQ page published in the top-ranked journal increases salary by 0.17% ($151.36 in 2006 dollars), so the above-cited 5-AEQ page increase from having a star advisor at a Tier 2 program as opposed to an unranked advisor at a Tier 1 program likely results in tangible monetary rewards.

ABBREVIATIONS

MIT: Massachusetts Institute of Technology

NRC: National Research Council

NYU: New York University

OLS: Ordinary Least Squares

UC: University of California

REFERENCES

Attiyeh, G., and R. Attiyeh. "Testing for Bias in Graduate School Admissions." Journal of Human Resources. 32, 1997. 524-48.

Bauwens, L. "A New Method to Rank University Research and Researchers in Economics in Belgium." CORE Discussion Paper, 1998.

Becker, G. S. Human Capital. 1st ed. New York: National Bureau of Economic Research, 1964.

Bratsberg, B., J. F. Ragan, and J. T. Warren. "Negative Returns to Seniority: New Evidence in Academic Markets." Industrial and Labor Relations Review, 56, 2003, 306-23.

Buchmueller, T. C., J. Dominitz, and W. L. Hansen. "Graduate Training and the Early Career Productivity of Ph.D. Economists." Economics of Education Review, 18, 1999, 65-77.

Cameron, A. C., and P. K. Trivedi. Regression Analysis and Count Data. Cambridge: Cambridge University Press, 1998.

Carson, R., and P. Navarro. "A Seller's (and Buyer's) Guide to the Job Market for Beginning Academic Economists." Journal of Economic Perspectives, 2, 1988, 137-48.

Collins, J. T., R. G. Cox, and V. Stango. "The Publishing Patterns of Recent Economics Ph.D. Recipients." Economic Inquiry, 38, 2000, 358-67.

Coupe, T. "Basic Statistics for Economists." Unpublished manuscript, 2001.

--. "Revealed Performances: Worldwide Rankings of Economists and Economics Departments, 1990 2000." Journal of the European Economic Association, 1, 2003, 1309-45.

Creel, M. D., and J. B. Loomis. "Theoretical and Empirical Advantages of Truncated Count Data Estimators for Analysis of Deer Hunting in California." American Journal of Agricultural Economics, 72, 1990, 431-41.

Crosta, P. M., and I. G. Packman. "Faculty Productivity in Supervising Doctoral Students' Dissertations at Cornell." Economics of Education Review, 24, 2005, 55-65.

Cushing, M. J., and M. G. McGarvey. "Sample Selection in Models of Academic Performance." Economic Inquiry, 42, 2004, 319-22.

Davis, J. C., J. H. Huston, and D. M. Patterson. "The Scholarly Output of Economists: A Description of Publishing Patterns." Atlantic Economic Journal, 29, 2001, 341-49.

Ehrenberg, R. G., and P. J. Hurst. "The 1995 Ratings of Doctoral Programs: A Hedonic Model." Economics of Education Review, 17, 1998, 137-48.

Ellison, G. "The Slowdown of the Economics Publishing Process." Journal of Political Economy, 110, 2003, 947-93.

Hirsch, B. T., R. Austin, J. Brooks, and J. B. Moore. "Economics Departmental Rankings: Comment." American Economic Review, 74, 1984, 822-26.

Hogan, T. D. "Faculty Research Activity and the Quality of Graduate Training." Journal of Human Resources, 16, 1981, 400-15.

Kalaitzidakis, P., T. P. Mamuneas, and T. Stengos. "European Economics: An Analysis Based on Publications in the Core Journals." European Economic Review, 43, 1999, 1150-68.

Kingston, P. and J. Smart. "The Economic Payoff to Prestigious Colleges," in The High Status Track: Studies of Elite Private Schools and Stratification, edited by P. Kingston and L. S. Lewis. Albany, NY: SUNY, 1990.

Krueger, A. B., and S. Wu. "Forecasting Job Placements of Economics Graduate Students." Journal of Economic Education, 31, 2000, 81-94.

Laband, D. N., and M. Piette. "The Relative Impact of Economics Journals." Journal of Economic Literature, 32, 1994, 640-66.

Liebowitz, S. J., and J. P. Palmer. "Assessing the Relative Impacts of Economic Journals." Journal of Economic Literature, 22, 1984, 77-88.

Long, J. S. "Productivity and Academic Position in the Scientific Career." American Sociological Review, 43, 1978, 889-908.

Moore, W. J., R. J. Newman, and G. K. Turnbull. "Do Academic Salaries Decline with Seniority?" Journal of Labor Economics, 16, 1998, 352-66.

Pieper, P. J., and R. A. Willis. "The Doctoral Origins of Economics Faculty and the Education of New Economics Doctorates." Journal of Economic Education, 30, 1999, 80-8.

Sauer, R. "Estimates of the Returns to Quality and Coauthorship in Economic Academia." Journal of Political Economy, 96, 1988, 855-66.

Scott, L. C., and P. M. Mitias. "Trends in Rankings of Economics Departments in the U.S.: An Update." Economic Inquiry, 34, 1996, 378-400.

Siegfried, J. J., and W. Stock. "So You Want to Earn a Ph.D. in Economics? How Long Do You Think It Will Take?" Journal of Human Resources, 36, 2001, 364-78.

Smyth, D. J. "The Determinants of the Reputations of Economics Departments: Pages Published, Citations and the Andy Rooney Effect." American Economist, 43, 1999, 49-58.

Spence, A. M. "Job Market Signaling." Quarterly Journal of Economics, 87, 1973, 355-74.

Stock, W. A., and R. M. Alston. "Effect of Graduate-Program Rank on Success in the Job Market." Journal of Economic Education, 31, 2000, 389-401.

Thursby, J. G. "What Do We Say about Ourselves and What Does It Mean? Yet Another Look at Economics Department Research." Journal of Economic Literature, 38, 2000, 383-404.

van Ours, J. C., and G. Ridder. "Fast Track or Failure: A Study of the Graduation and Dropout Rates of Ph.D. Students in Economics." Economics of Education Review, 22, 2003, 157-66.

Wu, S. "Recent Publishing Trends at the AER, JPE, and QJE." Unpublished manuscript, Hamilton College, 2004.

(1.) While the language used in much of the previous research leads us to infer that the authors were implicitly making this assumption, we note that such a stringent assumption is not required for the prediction to hold. Specifically, even if it is not the case that "only the very best students" enroll in elite programs, the graduates of such programs will still be predicted to be the most productive if those programs are the best trainers of future academic economists.

(2.) We realize that faculty might choose to take on advisees for reasons other than their potential research ability. Unfortunately, it is not possible to control for such unobserved factors.

(3.) We understand that observed differences in student-advisor matches are affected by cross-program differences in the supply of faculty with differing levels of research prominence. As such, the observed differences relating to advisor research prominence will to some extent be picking up the "faculty quality" element of cross-program quality differences.

(4.) As a specific example, Krueger and Wu (2000) report that 344 students applied to a given top 5 economics program in 1989, with 65 being admitted and 27 choosing to enroll.

(5.) We do note that there are a number of PhD recipients during the period from 1990 to 1994 for whom no dissertation advisor is specified. While small, the percentage is largest for 1990 graduates from the University of Chicago. The percentage missing decreases quickly over time and is practically nonexistent for the last 3 yr of our study period.

(6.) We estimated all models with smaller samples of years without significant differences in the results.

(7.) There is evidence of an increasing slowdown in the economics publication process. Specifically, Ellison (2003) finds that the average publication time had increased from 6 mo to over 2 yr.

(8.) Tier 1 programs are Harvard, Chicago, MIT, Stanford, Princeton, and Yale. Tier 2 programs are UC Berkeley, Pennsylvania, Northwestern, Minnesota, UCLA, Columbia, Michigan, Rochester, and Wisconsin. Tier 3 programs are UC San Diego, NYU, Cornell, Cal Tech, Maryland, Boston University, Duke, Brown, Virginia, North Carolina, University of Washington Seattle, Michigan State, Illinois, Washington University (St. Louis), and Iowa.

(9.) These metrics represent different methods for weighting publications according to journal quality, article length, authorship configuration, and article impact and include total number of articles and pages in all journals; total articles, adjusted articles, total pages, and adjusted pages according to Laband and Piette (1994); total articles in ten top journals (Kalaitzidakis, Mamuneas, and Stengos 1999); total articles with page size corrections in 24 top journals (Hirsch et al. 1984) and 36 top journals (Scott and Mitias 1996); and the methods for calculating an article's impact factor, based on citations, presented in Laband and Piene (1994) and Bauwens (1998).

(10.) According to author affiliation statistics in Wu (2004), among the 25 programs that publish more than 1% of all articles in the American Economic Review, 22 are top-ranked economics programs and 3 are members of the Federal Reserve System.

(11.) According to Coupe (2003), between 1969 and 2000, close to 131,000 individuals contributed articles to the economics literature. Among these, 71,983 contributed only one article, while 1,230 contributed ten or more articles with the maximum for one individual being 238.

(12.) A natural question might be what are the non-advising 659 of the top 1,000 economists doing? Among Coupe's 1990-2000 top 1,000 economic publishers, 486 were affiliated with top 30 U.S. economics programs, 145 with other U.S. PhD-granting economics departments, 45 with other U.S. academic programs, 2 with U.S. agricultural economics programs, 52 with foreign academic programs, 40 with the U.S. Federal Reserve System, 14 with the World Bank or International Monetary Fund, 7 with U.S. government agencies, 6 with think tanks, and 5 with private firms.

(13.) One notable difference between our approach and that of Buchmueller, Dominitz, and Hansen (1999) is that we estimate our productivity functions using a negative binomial specification, while they estimate theirs using ordinary least squares (OLSt. Buchmueller, Dominitz, and Hansen (1999) motivate their approach based on the fact that OLS is a more general specification. An important factor in our choice is that we wish to eventually calculate predicted productivity values, and in their comparison of out-of-sample predictions, Creel and Loomis (1990) find that "count data models predict substantially better than do OLS." Nonetheless, we note that as with Buchmueller, Dominitz, and Hansen (1999), we estimated our results both ways and "the sign and significance of key parameter estimates did not differ substantially across specifications."

(14.) The log-likelihood statistic for total articles, top 36 articles, and AEQ pages are 374.08, 294.22, and 156.58, respectively. Each of these values is above the tabled chi-square value of 11.35 at a significance level of .01.

MICHAEL J. HILMER and CHRISTIANA E. HILMER *

* We would like to thank Ron Ehrenberg, Dan Goldhaber, Kangoh Lee, Jayson L. Lusk, James F. Ragan, Jr., Mark Showalter, and seminar participants at Virginia Tech for helpful comments on previous drafts.

M. J. Hilmer: Assistant Professor. Department of Economics, San Diego State University, San Diego, CA 92182-4485. Phone 1-619-594-5662, Fax 1-619-594-5062, E-mail mhilmer@mail.sdsu.edu

C. E. Hilmer: Assistant Professor, Department of Economics, San Diego State University, San Diego, CA 92182-4485. Phone 1-619-594-5860, Fax 1-619-594-5062, E-mail chilmer@mail.sdsu.edu

TABLE 1
Summary Distribution of Students and Advisors

(a) Students and Advisors by Program Tier

 Tier 1 Tier 2 Tier 3

Student observations 524 743 625
Percentage of students 0.277 0.393 0.330
Total advisors 190 272 279
Percentage of advisors 0.653 0.465 0.323
 ranked in top 1,000

(b) Students and Advisors by Advisor Ranking Groups

 Lower
 Star Ranked Unranked

Student observations 536 573 783
Percentage of all 0.283 0.303 0.414
 students
Total advisors 132 213 396
Percentage of all 0.178 0.287 0.534
 advisors

(c) Students across Advisor Rank and Program Tiers

 Program Tier

Advisor Rank Tier 1 Tier 2 Tier 3

Star 0.500 0.219 0.178
Lower ranked 0.269 0.366 0.256
Unranked 0.231 0.415 0.566

TABLE 2
Distribution of Advisees across Advisors

 All Advisors Ranked Advisors

Number of Total Total Total Total
Advisees Observations Percentage Observations Percentage

1 343 46.29 105 30.61
2 142 19.16 66 46.48
3 91 12.28 51 56.04
4 57 7.69 40 70.18
5 37 4.99 27 72.97
6 28 3.78 19 67.86
7 6 0.81 6 100.00
8 15 2.02 11 73.33
9 6 0.81 4 66.67
10 5 0.67 3 60.00
11 2 0.27 1 50.00
12 3 0.40 3 100.00
13 -- -- -- --
14 -- -- -- --
15 2 0.27 2 100.00
16 -- -- -- --
17 -- -- -- --
18 1 0.13 1 100.00
19 1 0.13 1 100.00
20 1 0.13 1 100.00
21 1 0.13 -- --
Total 741 -- 341 46.02

TABLE 3
Summary Research Productivity by Program and Advisor Rank

 Publish Any Publish Top
 Articles 36 Articles Total Articles

All students 0.681 0.394 4.119 (5.486)

Program rank
 Tier 1 0.781 0.542 5.489 (6.279)
 Tier 2 0.661 0.376 3.795 (5.059)
 Tier 3 0.622 0.293 3.355 (5.043)

Advisor rank
 Star 0.778 0.560 5.481 (6.163)
 Lower ranked 0.675 0.428 4.347 (5.644)
 Unranked 0.619 0.257 3.019 (4.578)

 Top 36 Articles AEQ Pages

All students 1.295 (2.483) 13.730 (27.854)

Program rank
 Tier 1 2.216 (3.415) 24.899 (38.806)
 Tier 2 1.073 (2.046) 11.098 (22.871)
 Tier 3 0.787 (1.703) 7.494 (17.749)

Advisor rank
 Star 2.118 (3.066) 22.794 (34.508)
 Lower ranked 1.464 (2.681) 15.680 (30.743)
 Unranked 0.608 (1.503) 6.098 (15.896)

TABLE 4
Marginal Effects for Negative Binomial Regressions Controlling
for Advisor Rank and Program Tier

 Total Articles Top 36 Articles

Advisor rank
 Star 1.6710 ** (0.4074) 1.1688 ** (0.1948)
 Lower ranked 1.2262 ** (0.3358) 0.9758 ** (0.1586)

Program rank
 Tier 1 1.6963 ** (0.4097) 1.0623 ** (0.1899)
 Tier 2 0.4664 (0.3006) 0.2866 (0.1183)

First-job type
 Research position -- --

Individual characteristics
 Years since PhD 0.3541 ** (0.0891) 0.0854 ** (0.0322)
 International student -0.8741 ** (0.2537) -0.3370 ** (0.0926)
 Male 1.0320 ** (0.2702) 0.2536 ** (0.1007)

Advisor time constraint
 Other students supervised 0.0000 (0.0317) -0.0002 (0.0113)

Log likelihood -4,616.66 -2,652.62

[R.sup.2] .0151 .0424

[alpha] 1.5903 (0.0691) 2.3965 (0.1492)

 AEQ Pages Total Articles

Advisor rank
 Star 13.2519 ** (3.3032) 1.4506 ** (0.3117)
 Lower ranked 11.9321 ** (2.7052) 0.6482 ** (0.2448)

Program rank
 Tier 1 14.6468 ** (3.5038) 0.8336 ** (0.2866)
 Tier 2 4.1986 (1.8942) 0.0836 (0.2253)

First-job type
 Research position -- 4.1040 ** (0.2231)

Individual characteristics
 Years since PhD 0.6456 (0.5179) 0.2086 ** (0.0676)
 International student -2.4872 * (1.4603) -0.4962 ** (0.1939)
 Male 1.5332 (1.6575) 0.9362 ** (0.2051)

Advisor time constraint
 Other students supervised -0.0194 (0.1873) -0.0114 (0.0242)

Log likelihood -4,805.17 -4,429.62

[R.sup.2] .0122 .0550

[alpha] 7.7999 (0.3411) 1.1692 (0.0560)

 Top 36 Articles AEQ Pages

Advisor rank
 Star 0.8650 ** (0.1346) 9.6845 ** (2.3493)
 Lower ranked 0.6214 ** (0.1060) 6.7381 ** (1.7160)

Program rank
 Tier 1 0.5270 ** (0.1152) 6.6990 ** (2.0257)
 Tier 2 0.0991 (0.0817) 1.7346 (1.2750)

First-job type
 Research position 1.3011 ** (0.0814) 14.2184 ** (1.4755)

Individual characteristics
 Years since PhD 0.0376 * (0.0226) 0.3165 (0.3583)
 International student -0.1661 ** (0.0654) -1.9280 ** (1.0659)
 Male 0.2203 ** (0.0698) 1.3361 * (1.1726)

Advisor time constraint
 Other students supervised -0.0040 (0.0079) -0.0897 (0.1302)

Log likelihood -2,505.51 -4,731.88

[R.sup.2] .0955 .0273

[alpha] 1.6147 (0.1125) 6.7764 (0.3044)

Notes: Entries listed in the column heading are the dependent
variables. Regressions also include dummy variables indicating the
field in which the student's dissertation was written.

** Significant at 5% level; * significant at 10% level.

TABLE 5
Predicted Differences in Research Productivity Measures

 Total Top 36
 Articles Articles AEQ Pages

Star, Tier 1 9.326 (0.237) 3.267 (0.336) 34.691 (0.509)
Star, Tier 2 8.576 (0.233) 2.839 (0.329) 29.727 (0.493)
Lower ranked, Tier 1 8.523 (0.240) 3.023 (0.342) 31.745 (0.522)
Star, Tier 3 8.492 (0.229) 2.740 (0.321) 27.992 (0.493)
Unranked, Tier 1 7.875 (0.238) 2.402 (0.336) 25.007 (0.518)
Lower ranked, Tier 2 7.773 (0.231) 2.595 (0.327) 26.781 (0.487)
Lower ranked, Tier 3 7.690 (0.226) 2.496 (0.320) 25.046 (0.489)
Unranked, Tier 3 7.041 (0.221) 1.875 (0.310) 18.308 (0.480)
Unranked, Tier 2 7.125 (0.227) 1.974 (0.320) 20.043 (0.483)