Forecasting Demand and Supply of Doctoral Scientists and Engineers: Report of a Workshop on Methodology (2000)

It is important to the nation that there be an adequate number of scientists and engineers. Industries that rely on scientific and technological research and development are increasingly important in both the global and the American economies. If there are too few scientists and engineers, the economy and its competitive position, both now and in the future, are put at risk. An adequate supply of doctoral scientists and engineers is also important to the nation 's colleges and universities, since these institutions train both graduate students and undergraduates and carry out university research.

Conversely, if too many people trained as scientists and engineers cannot find work related to their training, costs also result to both the scientists and engineers, and to the federal government, state governments, and universities that have subsidized the many years of education that cannot be used appropriately. In the late 1980s, predictions that shortages of doctorates would emerge in the 1990s appeared not only in the technical literature (Bowen and Sosa, 1989; NSF, 1989), but also in the media (Sovern, 1989). Young people who went to graduate school at that time and expected a welcoming job market when they received their degrees were, in many cases, sorely disappointed.

A simple illustration of the way that economists analyze the costs of shortages or surpluses is shown in the four boxes in Figure 2-1.

The need to anticipate future demand for doctoral scientists and engineers, and to adjust factors that affect the supply to meet the demand, has encouraged the construction of forecasting models. Yet, as Johnson (1998) pointed out in his workshop paper, forecasting these markets beyond the very short term is extraordinarily difficult:

The future paths of levels of employment and wages in professional labor markets depend on the paths of at least ten exogenous variables. . . . Two or three of these exogenous variables, dealing primarily with the age structure of the population, can be predicted quite well for twenty or more years into the future. Other important exogenous variables, like the distribution of preferences of future college students for scientific versus other careers and future technological changes that might affect the demand for scientific personnel are inherently unpredictable. Further, there are important substantive issues involving, among other things, the heterogeneity of labor input within scientific fields and the substitution among groups of inputs with different quality/qualifications. These factors could drastically affect market forecasts and are not satisfactorily understood.

Even apparently predictable demographic change can become unpredictable if immigration increases.

Essentially, good forecasts require a well-specified model that correctly reflects behavior, is based on good data as well as careful forecasts of variables determined outside the model, and is produced by parties with no vested interest in actions taken in response to the forecast. Although past models, which were exhaustively reviewed in Leslie and Oaxaca (1993), often satisfied

one of these requirements,¹ few have been formulated to reflect the interaction of wages, quantities, and quality, let alone what the committee called the “neglected margins” of interchangability of workers across fields, by degree level, quality, or nationality.

What happens to forecasts when models leave things out? Put simply, they can mislead. For example, the NSF models of the late 1980s did not anticipate a deep economic recession and its effect on state and federal education and R&D budgets. Moreover, they did not anticipate the end of the Cold War and its effect on defense spending, which translated into fewer jobs for scientists and engineers. They failed to anticipate the impact of legislation that abolished mandatory retirement, thereby postponing (but not eliminating) the expected retirement of doctoral faculty who had been hired in the late 1960s and early 1970s to teach baby boom children. All of these unanticipated exogenous events worked to dampen the demographically based forecasts of an increase in demand for doctoral scientists and engineers. In addition, the models failed to account for the market mechanisms that operate to bring supply and demand into balance: wage adjustments, immigration, transfer of workers across fields, and adjustments in qualifications. The difference between the NSF forecast of the number of jobs that would be available for new Ph.D.s and the actual number of new Ph.D.s placed in science and engineering jobs is shown in Figure 2-2.

This incorrect forecast resulted from unanticipated changes in the economic environment, exacerbated by the model's neglect of market adjustment mechanisms. As Johnson points out, it is beyond the capacity of a long-term market forecasting model to

FIGURE 2-2 The NSF/Atkinson forecast of supply of new Ph.D. scientists and engineers in the early 1990s and what actually happened.

anticipate all the changes that can affect demand and supply. However, if regularly updated models had been available that took into account recent events and market adjustment mechanisms, and communicated the true limits of forecast accuracy, then forecast errors would have been smaller and less surprising. Such models could have provided an earlier warning that the market outlook would not be as rosy as the analysts of the late 1980s anticipated. Instead, policymakers were surprised by the increase in the numbers of new Ph.D.s who did not have definite job commitments upon graduation and by the rapid growth in the pool of postdoctoral students, especially in the biological sciences. In response, the National Science Board asked the National Research Council to study the question of how to “reshape” graduate education to

improve the chances that new Ph.D.s find appropriate employment outside of academia (NAS, 1993). At the same time, Massey and Goldman (1995) constructed a detailed, field-disaggregated simulation model that forecast considerable oversupply in most science and engineering fields. Their report pointed to the demand for graduate students driven by research funding as an important explanation of why production of new Ph.D.s in science and engineering exceeded academic demand. However, wages and other market adjustment mechanisms were also missing from the Massy-Goldman model.

Much of the discussion that followed the Johnson paper focused on how models might be improved. Areas for improvement fell into three broad categories:

• Data. Good data exist on Ph.D.s and especially on Ph.D.s in academia. Data on career paths of Ph.D.s in industry (or the classifications that would permit characterization of career paths) are lacking, as are sample sizes large enough to permit study of subpopulations in fine fields (e.g., bioinformatics). Since Ph.D. data are collected for individuals, consistency with data collected by occupations is difficult to achieve. Wage data (but not data on total compensation) by occupation are available, but not by different degree levels. Data are collected for international students who receive degrees from U.S. universities, but retention rates are difficult to determine. Data on immigrants who enter the United States after earning their Ph.D. have been collected only recently. The unavailability of data makes it difficult to model the mobility of scientists and engineers in response to changing market conditions. Finally, very little data allow for the measurement of changes of quality.

• Model specification. Most models to date have been “gap” models that project demand and supply separately. These models neglect wage adjustments and other market mechanisms that tend to modify demand and supply and bring them into

balance. Gap models thus describe structural conditions prior to the operation of these mechanisms or describe a world in which these mechanisms are believed to be ineffective. Better models would estimate demand and supply simultaneously and incorporate behavioral parameters so that changes in demand could, over time, feed back to supply and vice versa.

• Omitted variables. Most models leave out immigration (in part because good data are lacking) and workers in science with nonscience degrees. Barnow (1998) noted in his paper that for bachelor's degree holders, at least, these flows can be substantial. For those with a bachelor's degree in computer science, 53.5 percent take a job outside computer science, while only one-third of those with computer science jobs received their degree in that field. These flows may be less prevalent in markets for Ph.D.s, but are likely to become more important during periods of increased demand for particular types of workers. A current example is the field of bioinformatics.