Author’s note: Pay equity and pay equity analyses are a complex and nuanced issue, and context matters. Employers should always consult with legal counsel prior to engaging in any proactive pay equity studies and conduct those analyses under the attorney-client privilege.
As part of their obligation to "evaluate compensation system(s) to determine whether there are gender-, race-, or ethnicity-based disparities" under Executive Order 11246, 1 many contractors utilize multiple regression techniques to evaluate pay equity within their workforces. Such in-depth, statistical studies can highlight patterns of disparities in pay between protected class subgroups 2 that require follow-up investigation. Follow-up research may highlight areas in which salary adjustments are warranted for particular employees 3, making it possible to reconcile disparities where justified. In such cases, an employer can also consider a root cause analysis to understand whether disparities resulted from one or more specific compensation policies or practices that can be isolated.
Modeling legitimate explanations for pay differences
There are many advantages to the application of multiple regression techniques in evaluating issues of pay equity. Given the availability of compensation-related data for a group of similarly situated employees (i.e. salary data and information about employees’ standing on legitimate, non-discriminatory pay factors 4 ), employers can model the reasons why employees are paid the way they are paid. In fact, the statistical results serve as the initial step in a root cause analysis, as they provide information about (1) how well each factor explains differences in pay between employees 5 and how well the combination of factors explains differences in pay between employees.6 If the results of a multiple regression analysis indicate that the pay factors included in the model provide much less explanatory value than expected, it may be an initial indication that presumed compensation policies or practices are inconsistently implemented and/or that factors not included in the regression model are influencing pay.
Modeling disparities in pay associated with protected class status
Adding indicator variables of sex or race/ethnicity to the regression model allows an evaluation of differences in pay between protected class subgroups after controlling for the legitimate, non-discriminatory factors in the regression model. Three aspects of the regression results should be considered in evaluating the statistical disparity between two protected sub-groups:
- (1) Is the average difference in pay between the groups statistically significant? 7
- (2) How large is the average difference in pay between the groups?
- (3) Are there other legitimate, non-discriminatory factors excluded from the regression model that may account for the observed disparity?
If a statistically significant and practically meaningful 8 difference in pay between protected class subgroups cannot be explained by the legitimate, non-discriminatory factors included in the final regression model(s), employers may consider various adjustment strategies.9 A reactive adjustment strategy, however, could be enhanced by a root cause analysis of the policies or procedures that potentially produced the disparities but were not modeled in the initial regression analysis. 10
Because such a follow-up analysis can require significant time and resources, employers may be apt to conclude their annual pay equity analysis with the finalization of needed adjustments; however, failing to identify the root cause of pay disparities may predispose an employer to a recurring cycle of annual pay adjustments. For example, policies and practices related to setting starting salary and determining merit increases are potentially fruitful areas of exploration. 11
Subtle pay disparities between protected class subgroups can arise from practices serving legitimate, talent acquisition business interests. For instance, in their quest to attract top talent, employers may recruit employees from particular competitors or universities. In order to recruit the candidates from those "high potential" sources, employers may offer starting salaries that are higher in general than the starting salaries offered candidates internally promoted or recruited from less competitive sources. The presumption of the practice, of course, is that candidates from sought after recruitment sources have a more competitive set of knowledge, skills, abilities, and potential than other hires, and the compensation offered reflects that. Over time, if there is a higher proportion of a particular protected class subgroup hired from the "high potential" recruitment sources, compared to the proportion of overall hires, it may result in average starting salaries for one particular subgroup that are higher than for others. In such a case, the disparities could manifest over time as unexplained disparities in annual compensation studies. However, follow up analyses may support such an explanation.
Annual merit increase practices may also create unexplained disparities if not properly structured and monitored. Similar to starting salary, if one particular protected class subgroup receives higher merit increases on average, year-over-year, pay disparities could arise. However, pay disparities between protected class subgroups in merit increases are not problematic in and of themselves. If the processes used to make merit increase decisions are job-related, then existing disparities may be justified. Merit increase policies that provide guardrails around increases related to such factors as performance or range penetration reduce the likelihood that unexplained disparities in pay will arise.
Many employers use regression techniques to evaluate questions of pay equity and make adjustments to reconcile unexplained pay disparities between protected class groups. Without further exploring the root cause of pay disparities, however, employers might encounter year-over-year disparities that cannot be easily explained. Where the drivers of pay disparities are illuminated through a root cause analysis, it is potentially valuable for employers to evaluate whether the reasons can be defended as job-related (if feasible). Employing research methods, such as various validation strategies, when feasible, may allow employers clearer and more defensible justifications for such drivers. While such HR reviews and validation research require time and resources, their short term costs could result in long term benefits. Further, formalizing and standardizing aspects of the pay decision-making process can also help justify any differences in pay and help curb excess costs associated with compensation.
Aamodt, M., Cohen, D., Harpe, L., & Simpson, M. (2017). The science, art, and whimsy of making salary adjustments following a proactive pay equity study. Washington, DC: DCI Consulting Group, Inc.
Murphy, K. R., & Jacobs, R. R. (2012). Using effect size measures to reform the determination of adverse impact in equal employment litigation. Psychology, Public Policy, and Law, 18(3), 477-499.
Ricci v. Destefano (2009) 129 S. Ct. 2658.
1. 41 CFR 60–2.17(b)(3).
2. Such as between men and women or between different race/ethnicity groups.
3. Employers should be cautious in making adjustments to employee salaries without a strong basis in evidence (Ricci v. Destefano, 2009).
4. For example, information about how long each person has been with the organization.
5. Indicated by a combination of the size and statistical significance of correlations and regression weights.
6. Indicated by the model R-squared.
7. A statistically significant finding does not indicate that there is necessarily a problem with the compensation system. Statistical significance could arise because of situational circumstances that are unrelated to an employer’s pay policies.
8. Murphy & Jacobs (2012) discuss practical significance.
9. Aamodt, Cohen, Harpe, & Simpson (2017) describe pay adjustment strategies.
10. In any given compensation equity regression model, there are likely factors which influence pay that are excluded – either because the data are unavailable/unreliable or the factor is potentially tainted.
11. It is important to note that disparities in base pay may be a function of a variety of factors. The policies provided in the example above are not exhaustive, and chance variation is always a possible explanation.