Determinants of Research Production at Top US Universities

Specifically, the Nobel Prizes in physics, chemistry, physiology or medicine, and economics are considered. Fields Medals are awarded every four years to two to four mathematicians who are under the age of 40. “Staff” is defined as those who work at an institution at the time of winning the prize. Different weights are set according to the years in which prizes were won. The weight is 100% for winners in 2001–2003, 90% for winners in 1991–2000, 80% for winners in 1981–1990, 70% for winners in 1971–1980, and so on, and finally 10% for winners between 1911 and 1920. If a winner is affiliated with more than one institution, each institution is assigned a weight equal to the inverse of the number of institutions. For Nobel Prizes, if a prize is shared by more than one person, weights are set for winners according to their proportion of the prize.

Only 10–20% of the highly cited researchers work in “soft” science orientations, whereas almost all authors who publish in Nature and Science work in hard science fields. Among the awards being taken into consideration in the ranking, the majority are concerned with the “hard” sciences.

I also tested to control for the location of universities within states using a dummy for metropolitan areas with less than 500,000 inhabitants, another for areas with 500,000–2,000,000 inhabitants, and a dummy for more populated areas. These controls were never significant, and these results are not reported.

Note that I present the Huber–White standard errors that are robust to the presence of heteroskedasticity.

The null hypothesis of the weak instruments tests whose critical values are provided by Stock and Yogo (2005) is that the instruments are weak, that is, the relative bias of 2SLS estimation with respect to OLS regression is important. The critical values for the weak instrument test based on 2SLS bias for three instruments and one endogenous regressor are, respectively, for the 5% and 10% of maximal relative bias: 13.91 and 9.08 (see Stock and Yogo 2005)

The endogeneity test presented in the tables is performed in STATA 11 using the endog option after the command ivreg2. The null hypothesis of the test is that the endogenous regressor can be treated as exogenous. This test statistic is obtained by comparing two Sargan–Hansen statistics. The first is obtained by treating the suspect regressor as endogenous, and the second treats the suspect regressor as exogenous. This test is robust to the various violations of conditional homoskedasticity.

The S-estimator which is used in the estimations uses the Tukey Biweight function with a “tuning parameter” fixed at 1.547, Therefore, it resists up to 50% of outliers within the observations. The MS-estimator is a robust estimator which allows for both dummies and continuous variables. This estimator iterates alternatively an S- and an M-estimation.

One can be concerned by the fact that the outliers identified in the presented robust estimation could be different from those of an IV estimation. As a robustness check, I also estimate IV regressions where I apply the same identification technique at both stages of the IV estimation: to predict institutions’ revenue and their performances. I remain with significantly less observations, but this procedure ensures that the results cannot be driven by outliers. The results are even stronger. The corresponding tables as well as the first-stage estimations of the IV regressions are available in an online appendix.

The idea of the test is to compare the coefficients of an efficient and consistent estimator under the null hypothesis of no contamination with the coefficients of an estimator that is quasi-consistent under the hypothesis that contamination is present. If the coefficients are similar, the efficiency gain justifies the use of the efficient estimation. If they are too different, the robust estimation will be preferred.

With the other strategies discussed in the text, we could exclude up to 9 of the 15 first universities of the ranking.

In order to ensure that I get rid of the outliers, I also run weighted IV regressions where I give a weight of zero to the observations that are identified as outliers in the first or in the second step of the estimation. Results are not reported but remain very similar, though the significativity is lower (partly due to the fact that we exclude more observations).

The presentation of the results mainly rests on those presented in Tables 2 and 3.

As it can be seen from Figure A.2, Harvard, the MIT, the California Institute of Technology, Yale University, Stanford University, and the University of Chicago belong to these “bad leverage points”.

As explained above, I restrict the sample to institutions whose research expenditures are higher than 2% of their revenue and who offer doctorate.

I do not compare the magnitude of these coefficients as Wooldridge (2002) stresses that such comparison is not very informative.

To be more precise, all the “private” institutions present in the ranking are “private-non-for-profit”.