Seed Corn at the Millennium:

What Can We Learn from the Taulbee Data?

By Stu Zweben

As we approach the millennium, computer science finds itself in a position that bears great resemblance to that of the early to mid 80s. Enrollment in undergraduate computer science is booming. Many faculty positions in computer science departments reportedly are going unfilled, as are positions in industry (Freeman and Aspray 1999). Faculty are looking seriously at, and being aggressively recruited by, industry as an alternative to their current jobs. There is concern that we cannot keep enough students in our graduate programs to even keep the pipeline at its level of a couple of years ago, much less expand it to accommodate the increased demand for Ph.D. graduates. The forerunner of the CRA, in those days called the Computer Science Board, claimed circa 1980 that we were eating our seed corn by not being able to provide for the next generation of computer scientists in our educational institutions (Denning 1981). Do we have Seed Corn II upon us?

It would be useful to document such a situation as much as possible, so that appropriate actions can be taken to deal with it. In computer science, the best and most current source of data about human resources in our graduate departments comes from CRA's annual Taulbee Survey. What do these data tell us about the seed corn issue?

The Ph.D. Pipeline

The Taulbee Survey provides snapshots of the Ph.D. pipeline at three important points. The most often quoted figures are those at the end of the pipeline.

Figure 1

These are the numbers of Ph.D. graduates. But the survey also gathers information about the number of students passing the Ph.D. qualifying examination, and about the number of new Ph.D. students entering the academic departments.

Figure 1 shows the Taulbee data for each of these junctures since the mid 1980s, as reported by the Ph.D. granting computer science departments in the U.S. and Canada. The increasing number of Ph.D.s produced, as reported in the survey, is evident through 1994-95. This is a bit later than is often reported, and the discrepancy is due mainly to 1) the difference in number of departments reporting each year, and 2) the omission in our analysis of the computer engineering department data. To normalize for this variability in reporting, Figure 2 shows the pipeline data for the same period per department reporting. The number of departments reporting was taken from the Taulbee Survey's salary tables, though it is recognized that the number of departments reporting production, qualifier, and enrollment data may not be identical to the number reporting salary data. Again, the peak production occurs during 1994-95, at about eight per department. The last survey (for the 1998 graduating class) showed a value of 6.6 per department, a decrease of more than 15percent from the peak.

Figure 2

Figure 3 shows the differences in Ph.D. production based on departmental ranking. The Taulbee Survey categorizes the departments based on the ranking from the National Research Council. Prior to 1994-95, Canadian departments' figures were merged in these reports with those of the departments not ranked in the top 36. Also note that in 1995, the most recent NRC rankings were published, and this means that the actual departments included in each of the categories changed beginning with the 1994-95 survey.

From Figure 3, one can see that higher ranked departments produce on average considerably more Ph.D.s than do lower ranked departments. This figure also indicates that the decrease in Ph.D. production in the past two reporting periods has not been distributed uniformly across departments based on their NRC ranking.

At the Ph.D. qualifying exam level, Figure 2 shows a decreasing trend on a per department basis since the early 90s, with some recovery in 1996-97. Figure 4 shows the breakdown of the qualifier data by ranking. Here the decreasing trend is evident in all categories although, curiously, the collective departments ranked 13-24 seem to have recovered during the latest reporting period to their early 90s levels.

Figure 3

The number of new Ph.D. students per department declined from 1993-95, but for the last three years it has increased considerably, so it is now at its historic high of the late 80s. Figure 5 shows these data by ranking category. As can be seen from Figure 5, the recovery is most striking in the higher ranked departments.

Analysis

One might be encouraged by the increase in new Ph.D. students during the past three years, and predict that this will result in a significant increase in Ph.D. production in a few years. However, the qualifier data should cause one to be a lot more pessimistic about this. Suppose one assumes that, in general, a student will take the qualifier in the second year of study. This would mean that the qualifier trends should lag those of new Ph.D. students by a year. Figure 2 shows that, while not entirely true, this has been the basic pattern over the years. However, the gap in the two lines widened during the past couple of years. Thus, while the significant rise in new Ph.D. students per department has increased the number of those passing qualifiers per department, the proportions are going down. This is one piece of evidence in support of a seed corn problem; a smaller fraction of students appears to be entering the post-qualifier stage of study for the Ph.D. One cannot determine from these data if a smaller proportion of students are taking the exam, or a smaller proportion of those taking the exam are passing (or both). But the effect is worrisome.

Figure 4

The period between passing the qualifier exam and graduating with the Ph.D. is more variable. However, if the total time to the Ph.D. is about five to six years (the most recent Taulbee Survey showed the average to be just over five years), the lag between the qualifier trend and the Ph.D. production trend should be roughly four years. Figure 2 tends to confirm this hypothesis. Figure 6 shows the data from Figure 2 transformed so that the qualifier data is lagged one year from the new students data and the Ph.D. production data is lagged four years from the qualifier data.

The general decrease between 1990-91 and 1995-96 in the number per department passing the qualifier exam suggests that Ph.D. production per department will continue to show decline for the next couple of years, unless the proportion of those passing qualifiers who complete the degree increases from its historic levels. Of course, an argument can be made that the proportion of those passing qualifiers who go on to complete the degree is decreasing rather than increasing. Anecdotal evidence abounds of Ph.D. students leaving after passing the qualifying exam. It would be illuminating to see data at an intermediate point between the qualifier exam and the granting of the Ph.D. It is common for departments (or their institutions) to require a second intermediate hurdle, variously called by names such as the Candidacy or Preliminary Examination stage. Unfortunately, the Taulbee Survey doesn't collect data about this milestone, but such data may provide more conclusive evidence about the status of the Ph.D. pipeline.

Figure 5

The number of students graduating with the Ph.D. should not exceed the number passing the qualifier, unless there is a back door into the Ph.D. Figures 2 and 6 suggest, therefore, that the number of Ph.D.s produced per department will not rebound to its 1994-95 high in the foreseeable future.

Faculty Data

Production of new Ph.D.s is only one part of the seed corn issue. In order to educate the next generation of computing professionals, we need to currently attract and retain these persons within academic computing departments.

The number of advertisements for faculty positions is growing (CRN had to increase the size of its January issue to cope with the increased demand for advertising). If fewer new Ph.D.s are going into the academy, there will be a greater tendency for these faculty positions to go unfilled. If existing faculty are departing the academy with increasing frequency, this will damage our ability to not only teach new students but also perform vital research in the discipline and mentor the new faculty.

Figure 6

The Taulbee Survey collects data on the first jobs taken by new Ph.D.s, and on faculty departures. Last year, the survey also attempted to obtain information about the extent to which open positions are being filled, but the question was not worded satisfactorily. We are trying to improve our ability to collect this vacancy data.

Figure 7 shows, for new Ph.D. graduates whose first post-Ph.D. employment was known to be in the U.S. and Canada, the proportions of these graduates who went to industry and academia. Note that, within the "industry" category, the Taulbee data do not distinguish among positions in industry research, startups, development, etc. These data compare with NSF data that indicate, of new Ph.D.s in 1994 with plans for U.S. employment, 43% planned to go to industry and 49% to academia, while in 1996 these respective percentages were 50% and 44%. The data in Figure 7 illustrate the recent tendency for new Ph.D.s to go to industry with greater frequency than they do to academia. While industry employment is not to be discouraged, the effect of this trend contributes to the seed corn problem.

Of those new Ph.D.s who go into academia, the vast majority go to other Ph.D. granting computing departments. Figure 7 also illustrates this, and in fact shows a decreasing proportion who go to other than Ph.D. granting CS/CE departments. While in the early 90s there were about one-third going to such departments, now there are only around 10%. This does not bode well for these departments, which also are facing increasing enrollments and increasing demands for instructional personnel.

Figure 7

The Taulbee data on faculty departures is shown in Figure 8. On a per-department basis, the recent trend seems upward after some decline in the early 90s. The numbers are small enough so that it does not seem as if one can sound an alarm, though the trend is in a direction consistent with a seed corn problem.

Conclusion

The Taulbee Survey provides some evidence that there may be a seed corn problem, though the evidence to date is weak. There is a trend of a decreasing number of students per Ph.D. granting department who are continuing past the qualifying exam level, despite a recent upsurge in the number of new Ph.D. students per department. This likely will mean that the number of Ph.D.s per department over the next several years will remain below its historic high of just four years ago. There also are considerably more new Ph.D.s going to industry than to academia, and somewhat of an increase in the number of faculty per department who leave for non-academic positions.

Many persons intuitively feel that the seed corn problem is far greater than this. Additional data about the Ph.D. pipeline, such as at an intermediate point in the Ph.D. study, and data about vacancies and their ability to be filled, can improve our understanding of this problem. While we may not be able to collect this data in time to properly analyze the current situation, it should prove valuable for the future.

Figure 8

The vacancy problem goes well beyond those departments surveyed through Taulbee, since most of the undergraduate and masters computer science students are taught in departments that do not grant Ph.D.s. The decline in the number of new Ph.D.s who take positions in these departments is quite alarming.

Acknowledgments

My thanks to Bill Aspray and Stacy Cholewinski for their assistance in compiling these data and helping me understand how they were collected and how they might be interpreted. However, any errors in the arguments or their associated data should be attributable to me.

References

P. Denning, ed., "A Discipline in Crisis," Communications of the ACM, 24, 6 (June 1981), pp. 370-374.

P. Freeman and W. Aspray, The Supply of Information Technology Workers in the United States, Computing Research Association, 1999.

Stu Zweben is Professor and Chair in the Department of Computer and Information Science at Ohio State University. He is currently the chair of the Computing Research Association Surveys Committee. He is past Co-chair of the CRA Taulbee Survey Committee. He can be reached for comments at zweben@cis.ohio-state.edu.