Study of Salary Differentials by Gender and Discipline

ABSTRACT

Although it is 45 years since legislation made gender discrimination on university campuses illegal, salary inequities continue to exist today. The seminal work in studying the existence of salary inequities is that of the American Association of University Professors (AAUP), by Scott (1977) and Gray (1980). Subsequently, innumerable analyses based on versions of their multiple regression model have been published. Salary is the dependent variable and is modeled to depend on various independent predictor variables such as years employed. Often, indicator terms, for gender and/or discipline are included in the model as independent predicator variables. Unfortunately, many of these studies are not well grounded in basic statistical science. The most glaring omission is the failure to include indicator by predictor interaction terms in the model when required. The present work draws attention to the broader implications of using these models incorrectly, and the difficulties that ensue when they are not built on an appropriate sound statistical framework. Another issue surrounds the inclusion of “tainted” predictor variables that are themselves gender-biased, the most contentious being the (intuitive) choice of rank. Therefore, a brief look at this issue is included; unfortunately, it is shown that rank still today seems to persist as a tainted variable.

KEYWORDS:

• Model selection
• Regression
• Variable selection

Notes

1 This is easily tested in many regression packages, for example, in SAS, add the instruction “test < list of predictor variables>;”  when using the “proc reg” regression procedure.

2 Use “test gender ×degree, gender×employ.”

3 “Compression” and “inversion” occur when starting salaries are increasing faster than are merit raise increases; for “compression,” the lower ranked faculty person still has a salary lower than the longer serving faculty but the gap is shrinking, while “inversion” occurs when the starting salary has surpassed the salary of longer serving current faculty.

4 The often-used “rule of thumb” is to have about five times as many observations as there are predictor variables in the model; some authors, for example, Neter, Wasserman, and Kutner (1992) suggested 6 to 10.

5 As an aside, we caution that it could be that the two “groups” in Figure 6 may refer to some individuals who have fewer years and another group who have served many years so that these apparent differences merely reflect differences in longevity. The regression diagnostics should inform as to whether these differences are because of a non normality merging or simply because of longevity differences.