Recent concerns about the ability of the US to maintain its scientific leadership due to lower educational attainment in science and engineering (S&E) have led to recommendations focused on improving K-12 STEM education and incentives for students to pursue S&E education at the undergraduate and graduate levels (NAP2007). However, while needed, the recommended measures are insufficient to meet the emerging demographic realities [NAP2010]. For example, from 2001 to 2008 women formed slightly over 50 percent of the US population [USCB2012], while women’s rate of PhD attainment in science, engineering and health (SEH) for the same period ranged from 25.3 to 30.2 percent (SDR). Underrepresented minorities in the U.S. (URM) grew from 26.1 percent in 2001 to 28.6 percent in 2008 [USCB2012], but only represented 5.3–7.1 percent of US PhDs in SEH for the same period [SDR]. In this presentation, we analyze preconditions to the lack of diversity among SEH PhDs in the U.S. by focusing on the population of fulltime SEH graduate students to understand how current growth rates by race and gender may develop, and reason about the impact of plausible policy changes on the ensuing developments.

Our objective is to forecast the growth of demographic cohorts of full time graduate students in public and private institutions, taking into account how the growth of each cohort influences the other, and the impact of funding sources. We use the Survey of Graduate Students and Postdoctorates in Science and Engineering[GSS] as data sources, and a machine learning approach to time series modeling (Darlington 1996) based on support vector regression (Shevade et al. 1999) to forecast dependent and independent variables in parallel. Our dependent variables describe rates of fulltime graduate students by race, gender and US citizenship; our independent variables include sources of funding (federal such as NIH, DOE, NSF, non-federal, and self-support). An evaluation of forecasts 15-year onward with Mean Absolute Percentage Error (MAPE) and Direction Accuracy (DAC) corroborates the robustness of the modeling approach (on average, MAPE is 6% and DAC 86%). We also simulate the impact of policy changes, such as reducing to a half the number of foreign students and cutting down federal funding by 25%, on growth rates of graduate students by race and gender, using overlay/intervention variables. The emerging paradigm provides a promising framework to analyze diversity in the scientific workforce.

References

• NAP-07 (2007) Rising Above the Gathering Storm: Energizing and Employing America for a Brighter Economic Future. The National Academies Press, Washington DC.
• NAP-10 (2010) Expanding Underrepresented Minority Participation: America's Science and Technology Talent at the Crossroads. The National Academies Press, Washington DC.
• SDR (2001-2008) NSF Survey of Doctoral Recipients. 2001-2008 public files, http://sestat.nsf.gov/datadownload/.
• USCB-12 (2012) U.S. Census Bureau, Statistical Abstract of the United States: 2012, Table 6.
• GSS, http://www.nsf.gov/statistics/srvygradpostdoc.
• Richard B. Darlington (1996) A Regression Approach to Time Series Analysis, http://www.psych.cornell.edu/Darlington/series/series0.htm.
• S.K. Shevade, S.S. Keerthi, C. Bhattacharyya, K.R.K. Murthy: Improvements to the SMO Algorithm for SVM Regression. In: IEEE Transactions on Neural Networks, 1999.