Multivariate Analysis

Tenure

To explore whether the observed differences between men and women in their success at receiving tenure were statistically significant, and whether some of the variables described above explained those differences, we fitted a general­ized linear model to the binary outcome indicating whether a tenure decision was positive or negative. We included various institutional and departmental attributes as explanatory variables in the model, and used the method of generalized estimat­ing equations (GEE) to account for a potential correlation among tenure decisions in the same departments within the same institutions. The explanatory variables included in the model were the following: discipline, gender of the tenure candi­date (the variable of interest), prestige of the department, whether the institution was public or private, whether the department allowed faculty to extend the tenure clock for reasons including the arrival of a child, the percentage of tenure-track assistant professors in the department who were female, the percentage of females among the entire faculty of the department, and various two-way interactions between the gender of the tenure candidate and other variables. We did not include in this model any variables that might reflect the productivity of individual faculty members. The reason for this was that the subset of cases with complete informa­tion for all variables was relatively small.[94]

Results from this analysis are difficult to interpret, at least with regard to gender. While women appeared to be slightly more likely to be promoted and tenured than men, the effect of gender on tenure decision must be interpreted cautiously. This is because the interaction between the gender of the candidate and the percentage of females in the tenure-track pool was also evident. Women appeared to be more likely to be promoted when there was a smaller percentage of females among tenure-track faculty. Therefore, the difference between women and men in their tenure success was more pronounced in departments with fewer women assistant professors. After accounting for all the avenues through which gender affects tenure, across all fields, 93 percent of women and 83 percent of men who were considered for tenure were successful.

Assistant professors (both male and female) were significantly more likely to receive tenure at public institutions, where 92 percent of those considered became tenured, than at private institutions, where 85 percent gained tenure (p = 0.029). The probability of gaining tenure was greater in departments of lower (p = 0.017) or medium (p = 0.073) prestige compared to those in the high­est prestige category.

Because the presence of the interaction between the gender of the candidate and the percentage of women among tenure-eligible faculty prevented us from
interpreting the impact of either gender or percentage of females among assistant professors on tenure decisions, we did not attempt to untangle other associations. Figures 5-1 and 5-2 show the estimated probability of a positive tenure decision for men and women as a function of the percentage of tenure-eligible faculty who are female and as the proportion of female faculty in the department. To compute the probabilities in Figure 5-1, we held all other factors constant. Similarly, to compute the probabilities in Figure 5-2, we held the percentage of women among tenure-eligible faculty constant at 10 percent (two outer curves) or at 50 percent (two inner curves) for men and women.

Discipline, stop-the-clock policies, and overall departmental size were not associated with the probability of a positive tenure decision for either male or female faculty.

As a final comment, we note that when an interaction between a discrete covariate and other covariates in the model is present and the outcome variable is discrete (as is the case in our logistic regression model for tenure decision), unequal residual variances in each of the levels of the discrete covariate can have a profound effect on inference. Unequal group variances inflate the size of the estimated regression coefficients, thus introducing a bias in predictions relevant to differential outcomes for men and women. Therefore, trying to determine the effect of a covariate on, for example, male and female faculty cannot be done in

Percent

100

the usual manner. The class of models known as heterogeneous discrete choice models (e. g., Alvarez and Brehm, 1995[95]) has been proposed for analysis of this type of data.