Table 1.
Institution-level population summary statistics.
Table 2.
Professor-level sample summary statistics—Professional and socio-demographic.
Table 3.
Professor-level sample summary statistics—Nature of research.
Fig 1.
Individual and household earnings across fields.
Note: Based on 4,236 observations reporting both own and household earnings; Panel (c) is based only on 3,474 professors who report having a partner in their household; the correlation is . All averages reported are post-shrinkage to adjust for differences in sub-sample sizes across fields. Panels (b) and (d) includes estimates from the US Census Bureau [56].
Fig 2.
Earnings per institutions, ranks, tasks, and sources.
Note: Based on 4,388 observations. Observational earnings regressions are of the form: , where Yi is professor i’s earnings and
is the vector of either (Panel b) the professor’s hours spent on each tasks indexed by j, or (Panel c) the percentage points of the professor’s income due to each source indexed by j; the bar graphs in Panels (b–c) report the resulting estimates of the
coefficients; stars indicate statistical significance: p<0.1, ** p<0.05, *** p<0.01. In Panel (c), since the source percentages sum to 1, the base salary category is set to be the reference group. Institutional averages reported are post-shrinkage to adjust for differences in sub-sample sizes across institutions.
Fig 3.
Earnings per research metrics.
Note: Based on 3,323 observations matched to the publication/grant database. Observational earnings regression is of the form: , where Yi is professor i’s earnings and Xi is a variable or vector of their standardized inputs and outputs where f indexes fields; the bar graph in Panel (a) reports the estimates from univariate regressions including only one metric at a time; the regressions in Panel (b) are based on including all metrics (and other covariates); stars indicate statistical significance: p<0.1, ** p<0.05, *** p<0.01.
Fig 4.
Research productivity and time allocations.
Note: Based on 3,053 observations reporting non-zero research hours and matched to the publication database. Output is based on field-normalized citation-weighted publication rates per year or per research-hour. In Panel (c), stars indicate the significance of the unconditional correlation: p<0.1, ** p<0.05, *** p<0.01.
Fig 5.
Correlates of research risk-taking.
Note: Based on 4,186 observations reporting non-zero research hours. Panel (a) reports the results from lasso regressions predicting research risk with the full set of covariates from the sample; the stability selection share column reports the share of 100 bootstrap sub-samples that the covariate is selected in, and the bivariate correlation column reports that correlation for the full sample; p<0.1, ** p<0.05, *** p<0.01. Panels (b–d) show binned scatterplots relating researchers’ perceptions of the riskiness of their research (y-axis) to select covariates (x-axis) including field fixed effects to account for field-specific factors; the correlations illustrated in Panels (b–d) are ,
, and
, respectively.
Fig 6.
Nature of research, lifecycle, and tenure.
Note: Panels (a–c) are based on 4,095 observations reporting non-zero research hours and their age; Panel (d) is based on 2,180 observations from tenure-track professors at most 6 years before, or 15 years after, their tenure evaluation and includes only tenured professors post-evaluation. See the main text for details on how the intended output and audience scales are constructed.
Fig 7.
Correlates of the Bohr–Edison (“basic–applied”) spectrum.
Note: Based on 4,186 observations reporting non-zero research hours. Panel (a) reports results from the Principal Component Analysis (PCA) used to generate the Bohr-Edison score; more positive values indicate more applied, science, and more negative values indicate more basic, science. Panels (b-c) show field-level averages and earnings relationship, respectively. Panel (c) fits a quadratic line; the linear correlation is . Panel (d) reports the results from lasso regressions predicting the Bohr-Edison score with the full set of covariates from the sample; the stability selection share column reports the share of 100 bootstrap sub-samples that the covariate is selected in, and the bivariate correlation column reports that correlation for the full sample; stars indicate the significance: p<0.1, ** p<0.05, *** p<0.01.