*Names in bold indicate Presenter
To properly assess the size of these gaps as well as to optimally design policies targeting their elimination, accurate measures of the number of individuals majoring in STEM fields are essential. Most estimates of the number of STEM majors come from survey data. As STEM majors made up only 14% of all female bachelor degree holders and 32% of males in 2009, the entire population is relatively small. If one is interested in the male-female or black-white difference in a specific STEM area, the population shrinks even more dramatically. When using survey data to estimate such small population sizes, large errors are possible due to sampling variation and measurement error, making it extremely difficult to accurately measure gender and racial gaps. To illustrate, the 1993 NSCG estimates there are 2,717 women in the U.S. aged 35-45 who majored in civil engineering. This number should be approximately the same in the 2003 NSCG for the same group, now aged 45-55. Instead, it is 8,832, almost three times larger. In contrast, the corresponding estimates for men are 95,501 and 108,848, a difference of only 15%. This study remedies this problem by applying a Generalized Method of Moments (GMM) procedure (Black et al. 2012) to optimally combine data from multiple sources to produce considerably more accurate measures of the number of STEM graduates than what is possible from survey data alone. We use data from the NSCG, the American Community Survey (ACS), and Vital Statistics mortality records, producing precise measures of gender and racial gaps in over 50 different major fields within STEM in three different years (1993, 2003, and 2009). Our results allow us to identify which specific areas of STEM have the largest gender and racial gaps in college major attainment and to show the trend in these gaps over the last two decades. This information can inform policymakers about which specific areas of STEM to target when designing their policies to improve STEM education in the United States.
 Computed by the authors using the 2009 American Community Survey.