Discrimination in compensation can only be remedied by the periodic implementation of comprehensive studies within each post-secondary institution, using non-discriminatory measures in order to identify the internal inequities and equalize compensation.¹

- CAUT Policy Statement on the Human Right to Equitable Compensation

Compensation is more than salary, and can take various forms such as release time, research funding, size of the office, time to tenure and promotion, and workloads. There is little systematic data collection on these factors. Pay gap studies typically take a narrow view of compensation with a focus on salary differentials.

This guide looks at various approaches used to determine whether there are discriminatory salary differentials among academic staff. It does not provide deep detail into the analytical methods of studying pay gaps, but it does highlight some pertinent debates and issues.

A pay gap study may consider the broad array of reasons why women or another group in academia, on average, earn less than their male counterparts or another comparator group. Pay gap studies should take an intersectional approach to identifying possible discrimination in compensation.² Despite wide evidence of discriminatory compensation on the basis of race,³ for example, there have been few efforts in Canada to undertake pay equity studies beyond binary gender and even less taking multiple variables of discrimination into account.

The examples referenced in this guide therefore look at studies that examine how gender impacts pay. The approaches and promising practices, however, will be informative for pay gap studies looking at different and additional factors, including race.

These concepts are both critical when considering salary gaps.

Pay inequity is the difference in the salary between a dominant and an equity-seeking group with similar levels of experience and qualifications (they have the same rank, discipline, years of experience, years in academia, performance reviews, etc.). The difference in salary between the groups that cannot be explained through these variables can be understood as a direct form of pay discrimination. Most pay equity studies investigate primarily pay inequity.

Employment inequity refers to discrimination within the determinants of pay which impact the pay of equity-seeking academics. For example, equity-seeking group members may disproportionately face barriers in promotion to higher ranks, resulting in lower pay relative to dominant group. In this case, discrimination in employment equity is an indirect determinant of pay.  

Both pay and employment inequity may factor into the overall pay gap. For example, consider the Blinder-Oaxaca decomposition⁴ of the University of British Columbia (UBC)’s gender pay gap below, which shows the factors that make up the university’s pay gap.⁵ The majority of the gap can be attributed to factors that can be “explained”, such as rank, department, years at rank, and whether someone is a Canada Research Chair (CRC).  Employment inequity however may be influencing these factors, for example, less women in higher paid ranks or disciplines. The “unexplained” portion represents the part of the gap that cannot be explained by any other factor that is not gender; thus, it represents pay inequity.

Identifying pay inequity and the various forms of employment inequity requires different analytical approaches. Pay and employment inequities also require different remedies or solutions. It is important both of these elements of discrimination be taken into account when explaining the adjusted and unadjusted pay gaps.

The sections here will outline steps and considerations that associations can take when planning and undertaking a pay equity study. It will discuss the formation of a committee, planning and analysis, and data and research methodologies.

Academic staff participation in an equal compensation study is required in order to enhance understanding and acceptance of the process, goals and outcomes.

Many pay equity studies that are publicly available have been done under a joint faculty-management committee. Sometime the Committee has the study undertaken by faculty members, though sometimes an external consultant can be engaged to do the work. The consultant may be faculty from another institution or an independent consultant. There may be different joint committees as part of the study. For example, at Simon Fraser University, one committee was struck to undertake the analysis of faculty salary data, and a second was tasked with forming recommendations based on the findings of the analysis. 

A joint employer and association committee is best placed to undertake pay gap studies in order to have access to the necessary data, agree upon the methodology and ensure remedies and recommendations are implemented.

Associations have also undertaken their own studies, to bring pressure on administration to take action on inequities in compensation.  The study undertaken by the Association of Academic Staff of the University of Alberta, published in 2017, is an example of a study undertaken by a faculty-only committee.

In some cases, individual faculty have undertaken studies without formal committee representation. In these cases, the study may be published as an academic work,⁶ or a study undertaken by a faculty member which may be brought to the association to motivate action against inequities identified in the study.⁷

If the goal of a study is to seek a monetary remedy for found pay inequity, then the analysis will require a regression analysis of individual salary and employment data. If this data is not made available to the association or a joint Committee, one approach may be to plan the analysis in phases. The first phase can be exploratory, employing analytical methods that help the committee identify areas where there are potential pay or employment inequities.

In planning, the committee should carefully consider the scope of the analysis. One possible danger is setting too narrow of a scope for the analysis and coming out of the analysis with very few useful results. For example, this situation may arise if it is agreed to investigate only pay inequity and they have 1) a small faculty, and/or 2) a relatively small overall pay gap. In this case, there is greater likelihood that the pay discrimination found may be statistically insignificant.⁸ The statistical insignificance may have implications on the validity of the findings in a bargaining or arbitration context.

Many quantitative pay gap studies undertaken by Canadian universities have used administrative salary data from the university’s human resources departments. For some academic staff associations, their collective agreement stipulates data sharing agreements with the institution, where the administration gives salary data to the association every year. The association may maintain its own database of salary data for investigating pay anomalies or for regular tracking of the adjusted pay gap.

In some cases, it may be very difficult for associations to obtain salary data from administration. In one pay gap study, the University of Alberta faculty committee worked around its data issue by using the province’s “sunshine list” to compile its own salary data base. However, this approach required a large amount of manual data entry to create variables critical for pay inequality analysis.⁹

Another approach to obtaining salary data may be to use data from Statistics Canada’s Full-time University and College Academic Staff System (UCASS). UCASS collects institutional data on university salaries by gender, age, rank, discipline, years from degree and other variables that can inform salaries. CAUT has longitudinal UCASS data by institution available for use by academic staff associations.

For other variables, if the institution does not wish to disclose, a research plan can be created that digs deeper into the issues and potential drivers of the pay gap that are of interest. For example, a sample could be done with faculty that volunteer to participate to determine if there is a gap that requires further analysis. Data on faculty member’s self-identification can also be valuable for understanding inequities from an intersectional perspective.¹⁰ The University of Alberta faculty study used data on visible minority status to explore the potential interaction of gender and racial bias on pay.  

Some studies have also used survey data, which can be useful in investigating employment inequities. The University of British Columbia’s study of 2010 points out that “Working Climate” studies have been used by some universities to understand the systematic drivers of pay inequities, noting that salary correction mechanisms themselves generally do not prevent gender pay inequities from arising.¹¹

This section provides a general overview of some analytical methods that could be used in a pay gap analysis. Analytical methods are categorized as either exploratory and prescriptive, where exploratory methods may be useful for preliminary, “first round” of analysis. Prescriptive analysis allows the researcher to quantify the degree of gender inequity within a given facet of the pay gap, which may be useful for establishing a monetary remedy.

A thorough descriptive analysis of the administrative data can be an incredibly useful aspect of a pay gap study. While many insights from descriptive analysis could also be gleaned form other analytical techniques, descriptive analysis should not be overlooked. It can help the researcher identify potential mechanisms driving the overall pay gap (both pay and employment inequities), which can inform the planning of subsequent analysis and the specification of regressions. It can also help the researcher identify factors that drive gender pay disparities, allowing the committee to focus its analysis on areas that contribute largely to the pay gap.  

A descriptive analysis could examine counts of faculty, or mean and median salary through cross-tabulation of variables relevant to compensation: gender, rank, discipline/department, years of experience, age, etc. Sub-populations of the faculty can also be examined along a salary distribution, and distributions of men and women can be compared.

The method seeks to decompose the difference in average salaries of men and women to identify the drivers of the pay gap.

There are many variations on this method, but the general method requires the researcher to estimate two regressions where salary is regressed onto variables that capture the determinants of salary.¹² In one regression, only men are included, and the other uses only salary data of women. The results of the regressions are then compared to determine to what extent the gap is determined by the “explainable” factors, which are included in the regressions, or “unexplainable” factors which suggest the presence of pay inequity.

This method may be useful for understanding the contributors of the overall pay gap and their relative magnitude, allowing the committee to identify areas where potentially large employment inequities exist. A technical description of the method can be found in Brown and Troutt’s paper examining gender pay disparities at the University of Manitoba.¹³

This is the most common form of analysis in pay gap studies undertaken at Canadian universities. Here, the researcher typically uses Ordinary Least Squares (OLS) regression to create a statistical model that reflects the system of pay determination at the university. Using individual-level salary data, the researcher regresses salary¹⁴ onto variables that represent factors that determine pay, such as rank, department/discipline, age/years at rank/years of experience. Depending on data availability, the regression may also include other variables that may influence pay such as Canada Research Chair status, number of publications, etc. or other demographic data, such as racialized group membership.

Importantly, a variable denoting gender is also included in the regression. The resulting coefficient for this variable represents the average pay gap between men and women at the institution and is a measure of pay inequity. Some academic staff associations have received monetary remedies based on the magnitude of this coefficient; the cases of McMaster University,¹⁵ Wilfred Laurier University,¹⁶ University of British Columbia,¹⁷ and Simon Fraser University¹⁸ are some examples.

When specifying a regression, researchers should consider any variables that are important for the determination of pay at their university. Since different universities have different pay structures, there may also be differences in how models are specified. However, there tend to be core variables included in almost all salary models, namely gender, rank, discipline, and some measure of years of experience or age.

There are other regression and statistical techniques that may be useful for regression analysis of salaries.¹⁹ One such method is hierarchical linear modeling which is a form of OLS regression that accounts for the “clustering” or “nesting” of observations, which may be applicable in some instances given that faculty are organized by department/faculty. The analysis undertaken at the Wilfred Laurier University use this technique.²⁰

Regression of this type can be done on different types or components of salary, such as starting salaries, market differentials, retention payments, or stipends. The committee for the 2015 Simon Fraser University examined the components of an academic salary, which includes base salaries and “off-scale” amounts like market differentials and retention awards. From its analysis, the committee determined that gender pay disparities were driven by disparities in market differentials, not base salaries.²¹

Using individual-level salary data, researchers and committees can also investigate the degree to which determinants of salary, like rank, years of experience, or other “tainted” variables are impacted by gender.

In its 2010 study, the University of British Columbia undertook a Linear Probability Model/Probit Regression to determine whether gender impacted the probability of an academic being promoted to full professor rank. The variable for rank was regressed onto variables denoting experience, department, and Canada Research Chair and Distinguished University Professor status. Based on this analysis, gender was found to be a statistically significant determinant of being promoted.

Similarly, Simon Fraser University’s 2015 study used a Cox Proportional Hazard model to “identify whether or not gender is correlated with the time spent as an Associate before becoming a Full” by estimating the odds of promotion.²² The analysis did not find evidence that gender impacted the odds of promotion, but it did find that “[…] faculty who take medical leaves face greatly reduced odds of promotion to Associate and Full Professor. Further, Assistant Professors who take parental leaves face much lower odds of promotion to Associate Professor. These effects are similar for men and women, but women are much more intensive users of both leave types than are men.”²³

It is important to be aware of the justification for including and excluding variables. Fundamentally, the specification of the model should reflect the pay structure to which faculty are subject. If the analysis is being done within a committee, specification of the model may be discussed and negotiated between committee members.

There are a multitude of complexities in undertaking regression analysis which have spurred debate. Some of these issues are introduced below. When undertaking analysis, it may be useful to consider how these different perspectives in these debates would shape an analysis.

Variables that may be influenced by gender discrimination are considered “tainted”. For example, variables related to rank may be tainted if the rank of an academic is influenced by his/her gender. Including tainted variables may reduce the magnitude of the gender coefficient since the tainted variable is to some degree also a measure of gender.²⁴ Some may argue that tainted variables should not be included in a regression because they lead to an under-estimation of the impact of gender on pay, camouflaging the total impact of gender on salaries.

However, Dean and Clifton make the point that excluding tainted variables, specifically rank, may lead to the misattribution or remedies. If rank is not included in a salary regression but is a determinant of pay, the gender coefficient resulting for that regression suggests a financial remedy that “essentially amounts to paying a hypothetical average female as if she had been promoted at the same rate as the hypothetical average male.”²⁵

The problem of tainted variables can be seen as the intersection of pay and employment inequity and shows the importance of setting clear objectives prior to undertaking study. If the objective is to measure pay inequity, then tainted variables relevant to the determination of pay should be included (barring any collinearity problems). If the committee is seeking to determine the overall impact of gender on salaries (pay inequity plus employment inequity), it may wish to exclude tainted variables. The decision of whether to include or exclude tainted variables may depend on the function of the study.

Outliers are another important consideration since OLS regression can be very sensitive to extreme values in the dataset. There are several methods to determine the impact of outliers on results which will not be discussed in detail here. However, more generally there is some debate on how outliers should be treated in these studies. Some may argue that at many universities, “super-star” earners tend to be primarily men. Since this trend is a manifestation of the systematic pay difference of men and women, these observations should remain in the data.²⁶ However, one may argue that since OLS results reflect average impacts, extreme values should not be included because they are not representative of average salaries. Additionally, there are other regression methods that are less sensitive to outliers, such as robust regression or median regression, that may mitigate the impact of outliers on regression results.

Some argue that it is not necessary to consider the statistical significance of the gender coefficient in the context of a university pay equity study because the data used to estimate the model is a census, not a sample. The University of Lethbridge’s pay equity study of 2008 provides a more detailed explanation of this point.²⁷  However, some also argue that statistical significance is relevant because there may be imperfections with the data and/or with the regression specification. It is unlikely that regression is a literal replica of the system used to determine pay at the university. Testing for statistical significance acknowledges this discrepancy.²⁸

Ideally, the analysis undertaken by the committee or academic staff member(s) will support the development of specific recommendations. These recommendations may suggest monetary remedies that correct for past pay inequity, as well as systems for the ongoing monitoring of the adjusted pay gap to prevent/correct future pay inequity. These monitoring systems could be tied to anomaly funds that are specified within the academic staff association’s collective agreement. 

Some studies have also recommended policy change that help to mitigate employment inequity. Studies done by Simon Fraser University the University of British Columbia and McGill University, for example, combine analysis of both pay and employment inequity, with recommendations that address many facets of the pay gap.

¹ CAUT. Policy Statement on the Human Right to Equitable Compensation, November 2017.

² Woodhams, C., Lupton, B., and Cowling, M. "The Snowballing Penalty Effect: Multiple Disadvantage and Pay", British Journal of Management 26(1): 63–77, January 2015.

³ CAUT. Underrepresented and Underpaid: Diversity and Equity Among Canada’s Post-Secondary Education Teachers, April 2018.

Conference Board of Canada. Racial Wage Gap, 2020: Racial Wage Gap - Society Provincial Rankings - How Canada Performs (conferenceboard.ca)

Longhi, Simonetta and Malcolm Brynin. The Ethnicity Pay Gap. Institute for Social and Economic Research, University of Essex. Equality and Human Rights Commission, 2017.

⁴ Jann, Ben .Blinder-Oaxaca decomposition. Section 2.4.1. 2008. The Blinder–Oaxaca Decomposition for Linear Regression Models - Ben Jann, 2008 (sagepub.com)

⁵ Karen Bakker et al. An Analysis of the Gender Pay Gap in Professorial Salaries at UBC. University of British Columbia. 2010.

⁶ Faculty from the University of Manitoba have published two academic studies investigating the gender pay gap at the university:

Brown, L.K., Troutt, E.. Sex and Salaries at a Canadian University: The Song Remains the Same or the Times They Are a Changin'? Canadian Public Policy Vol. 3, Issue 3. 2017, Canadian Public Policy.

Brown, L.K., Troutt E. and Prentice, S.. Ten Years After: Sex and Salaries at a Canadian University. Canadian Public Policy. 2011.

⁷ In the case of Carleton University, faculty members gained access to salary data and undertook their own analysis. They then brought thee findings to the faculty association to motivate action to address the found inequities (Interview with Carleton faculty member, 2018).

⁸ If the committee chooses to do a regression analysis to or a given standard deviation and regression coefficient, statistical power increases with sample size.

⁹ Details on data collection for this study can be found at page 2 of the report:

Rosychuk, Rhonda J. et al.. Gaps in Professorial Compensation by Gender, Visible Minority, and Indigenous People at the University of Alberta. AASUA. 2017.

¹⁰ Importantly, there may be restrictions in linking survey data to salary data, as discussed by Michael Ornstein in his analysis of salary anomalies at the University of Windsor, 2002. 

¹¹ Karen Bakker et al.. An analysis of the Gender Pay Gap in Professorial Salaries at UBC: Report on the Pay Equity (Data) Working Group. University of British Colombia. 2010.

¹² Regression analysis is discussed in the section 2.4.2, “Regression analysis of salaries – pay inequity”.

¹³ Brown, L.K., Troutt, E.. Sex and Salaries at a Canadian University: The Song Remains the Same or the Times They Are a Changin'? Canadian Public Policy Vol. 3, Issue 3. 2017.

¹⁴ In some cases, the natural log of salary is also used.

¹⁵ Flaherty, Colleen. Leveling the Field. Inside Higher Ed. April 30, 2015.

¹⁶ Wilfrid Laurier University. Laurier gender-equity analysis results in salary increases for the university’s female associate and full professors. May 8, 2017.

¹⁷ Bradshaw, J.. UBC gives all female tenure-stream faculty a 2 per cent raise. The Globe and Mail. February 2, 2013.

¹⁸ Simon Fraser University Faculty Association. Salary Equity Agreement. December 19, 2016.

¹⁹ Some additional methods are discussed in the following publications:

Strathman, J. G.. Consistent estimation of faculty rank effects in academic salary models. Portland State University. 2000.

Johnson, C. B., Riggs, M. L., & Downey, R. G.. Fun with numbers: Alternative models for predicting salary levels. Research in Higher Education, 27(4), 349-362. 1987.

²⁰ Rutherford, J., Brunskill C., et al.. Final Analysis and recommendation of the Bi-Lateral Committee on Gender-based Pay. 2017.

²¹ Kessler A. and Pendakur K., Gender Disparity in Faculty Salaries at Simon Fraser University. Department of Economics, Simon Fraser University. 2015.

²² Ibid.

²³ Ibid.

²⁴ It may also introduce collinearity, compromising tests of statistical significance.

²⁵ Dean, J. & Clifton. R.. An evaluation of pay equity reports at five Canadian universities. Canadian Journal of Higher Education, 24(3), 87-114. 1994. 

²⁶ Interview with McMaster faculty member, 2018.

²⁷ Mellow, Muriel, et al.. Salary Equity Committee Report to the University of Lethbridge. 2008.

²⁸ It is also important to recognize the critical relationship between statistical significance and sample size. For a given pay gap, as sample size increases, statistical significance will also tend to increase. Thus, the issue of statistical significance is particularly relevant to universities with a relatively small number of faculty.