This law has several interesting contrasts with other EEO statutes, particularly Title VII:
See Gutman, Koppes, and Vodanovich (2011) for an overall discussion of ADEA, including how the law has evolved over the years with amendments and court decisions.
Recently the U.S. Court of Appeals for the Third Circuit rendered an opinion in Karlo (2017) that touches on adverse impact statistical analyses for age discrimination. The major thrust was a discussion regarding discrimination against “older, older” people in favor of younger people who are still members of the protected class. For example, the employer could have a practice that favored people under age 50 but whacked those 50 and above. The 40–49 folks are in the protected class. The courts have distinguished between age discrimination and “over 40” discrimination; ADEA is based on age.
This writer and others (e.g., Dunleavy and Morris, 2017, p. 9) took it as generally accepted practice that there could be dichotomies such as in the example above, not just under 40 versus 40 and above. The law is more complex. Whether the case is treatment or impact, and the location of the court, can matter.
Karlo and other plaintiffs were employed by Pittsburgh Glass Works (PGW), which manufactures automotive glass. There was a need for several reductions in force (RIFs) because of deteriorated sales during the Great Recession. Plaintiffs initially secured conditional certification for a collective action[i] for additional plaintiffs at least 50 years old when terminated. The case was then transferred to another district judge. On a motion from PGW the court decertified the collective action because the additional plaintiffs had claims factually dissimilar from the original plaintiffs. PGW then moved to exclude the plaintiffs’ two expert witnesses, Drs. Michael Campion and Anthony Greenwald. The motions were granted. Finally, PGW moved for summary judgment. The court granted this, noting that a 50-and-older group was not allowed under the ADEA and the exclusion of plaintiffs’ expert for adverse impact meant that this claim collapsed.[ii] Plaintiffs appealed the exclusion of their experts and summary judgment for PGW.
Both sides were backed by heavy-hitter organizations that filed amicus curiae (Latin, “friend of the court”) briefs: EEOC for the plaintiffs, and the U.S. Chamber of Commerce and the Equal Employment Advisory Council (an employer association) for the defendant.
The central question for the appellate court was whether subgroups of the 40-and-over protected class were allowed, here the 50-and-over group, for adverse impact analysis. In the court’s view, this could be answered in the affirmative based on the plain text of the statute, specifically the focus on age as the protected trait (and not membership in a protected class) and the rights of individuals[iii] (in contrast to the protected class as a whole). The Supreme Court’s decision in O’Connor (1996) clarified that ADEA proscribes age discrimination, not over-40 discrimination. Although that was a treatment case, the underlying issue for either treatment or impact regarding the nature of age is the same. The conclusion is reinforced by Teal (1982), which stands for protection against discrimination accruing to the individual. Discrimination against some members of the protected group while favoring others is still discrimination.
The court’s position is at variance with that of other circuits regarding impact cases. The Second (Lowe, 1997) and Eighth (EEOC v. McDonnell Douglas, 1999) do not allow for age groups other than at the 40 year breakpoint. The Sixth allows various groups for treatment (Barnes, 1990) but in a nonprecedential opinion, not for impact. Some lower courts in other circuits have been more flexible (Paetzold and Willborn, 2016). The issue seems to be the concern that formation of subgroups within the protected class is subject to manipulation; having multiple subgroups could impose on employers a potentially impossible task of ensuring parity for every subgroup that might be proposed. The Karlo court found arguments in the other circuits unpersuasive. Should a plaintiff try to gerrymander an age category specific to that person’s benefits, the courts can deal with it. The possibility of gerrymandering is no reason to negate the protection for older workers, even those older within the protected class.
The Karlo court recognized that age is a continuous variable, unlike the categories of race or sex. “On account of that difference, the statistical techniques common to Title VII cases are not perfectly transferrable to ADEA cases. If, for example, the comparison group in Teal omitted some black employees who took the written test, the statistics would have failed to address whether there was a disparate impact ‘because of … race.’ . . . But with the ADEA, by contrast, a comparison group that omits employees in their forties is fully capable of demonstrating disparate impact ‘because of … age’“(Karlo, p. 24, statute citations omitted).
The court’s statement is a bit confusing. What’s under discussion is not statistical methodology but how groups are formed for analysis. “Comparison group” appears to refer to the plaintiffs’ age group. In the case where the plaintiff group was those 50 and above, “those in their forties” were not omitted from the analysis. They were included in the under 50 group.
A more fundamental question is whether we should be concerned about age or individuals (their proportions in the outcomes), or both. Presumably both kinds of analyses are relevant.
So what might be transferrable or not regarding statistics? The court made frequent reference to Paetzold and Willborn’s (2016) book, so that will figure in this discussion.
Categorical analysis. Although Paetzold and Willborn deplore the use of categories with a continuous variable such as age, the dichotomies that serve us well with other protected classes seem to be standard procedure here. But dealing with a continuous variable invites arbitrary categorization within the protected class, chosen by the expert to produce what the client wants. What makes sense will depend on circumstances. Here, 50 and above was good. The court discounted the possibility of the octogenarians claiming discrimination because the folks in their seventies were allegedly favored. There likely would not be statistical power with the few people involved to get statistical significance. A generalization would be that that a floor age is workable if supported by the circumstances of the case, that is, the plaintiff group is at least x years and includes everyone x and over. A lower x implies more inclusiveness of the 40-and-above class. A higher x implies less statistical power, which is the brake on gerrymandering an extreme age category. Using both a floor and ceiling (e.g., no less than 50 and no greater than 52) likely invites scrutiny. Using just a ceiling seems a bit strange; to the extent that the older workers above the ceiling were favored, it may be hard to make a case.
Ironically, detecting rather than contriving age bands that drive statistical significance would seem to demand an examination of categories. For example, there is a statistically significant disparity in outcomes between the under-40 and 40-and-above people. But closer examination reveals that statistical significance is driven by disparity in the 50–54 band. The question could be whether it is plausible that discrimination affects just that band.
One possible analysis would be to run a chi square test across “reasonable” categories, such as 5-year bands. This could be accompanied by a test for linear trend to see if being older implies more likelihood of adverse action. Paetzold and Willborn (2016) mention the Cochran-Armitage test.
Age is not the only consideration open to categorization. For example, there may be an organization-wide RIF carried out according to centrally determined procedures. Plaintiffs may be inclined to note that commonality of procedure and include all employees in an organization-wide analysis. The employer might be more inclined to point out that the RIF was conducted by unit, in which the employees, decision makers, and need for reduction varied. This would argue for unit-based analyses with smaller numbers and thus less power to detect a true difference. Paetzold and Willborn (2016) note that a Mantel–Haenszel approach (stratification by unit, but with an overall statistical test) might be useful. There is also a Fisher method for analyzing across strata with continuous variables.
Continuous variable analysis. The most obvious analysis is to look at the mean age of those adversely impacted versus that of folks not impacted. The mean age of those whacked in a RIF might far above that of those who stayed. But watch out for outliers. If one of the employees was Methuselah (lived 969 years; Genesis 5:21-27) toward the end of his career, results might be skewed. Something similar might happen with less extremely aged, but more numerous, employees.
Likely no practitioner would do an analysis of people outside the plaintiff group for an impact case. However, with a “pattern or practice” treatment case that implies that discrimination was the standard operating procedure, it might be useful to show statistically that there was an age effect throughout the organization. For example, a RIF was conducted across all organizational units; layoffs were based on seniority; greater seniority is correlated with greater age; no new employees were hired; but the retained workforce comes out younger as measured by average age[v]. That would be the plaintiff’s analysis; the defendant might want to show that, assuming there were any age issues, they were limited and not standard operating procedure.
Paetzold and Willborn (2016) suggest the Wilcoxon rank test to avoid the need for arbitrary categorization. Use of ranks may also avoid the Methuselah problem. They also mention the Kolmogorov-Smirnoff test for differences between age distributions.
Logistic regression provides a method for examining the effect of age as well as other independent variables (continuous or categorical) on a dichotomous outcome. Biddle (2011) indicated that an analysis of independent variables other than the protected class generally serves no purpose for the plaintiff; the burden is just to establish the numerical disparity, not to explain it. Regression with age has the problem of the under-40 group (or other comparator) contributing to the regression weights; see Paetzold and Willborn (2016) for suggestions on recoding for more interpretable results.
Finally there is survival analysis and its relations. This is a variant on the regression theme notable for dealing with “censored” data, that is, people coming and going so that not all who are in the study have been there for the identical time period.
Age and the Peter Principle
The popularized Peter Principle holds that employees rise to their level of incompetence, that is, not so incompetent as to be removed but not promotable. This implies that less-than-stellar employees grow older as they stagnate in their positions. The application to age cases was discussed by Grossman et al. (2007). At the heart of the argument is that persons of different ages in (or applying to) the same jobs and units are not necessarily similarly situated.
Siskin and Schmidt (2017) noted that, for promotion, the usual regression analysis examines whether group differences remain after all factors that influence the likelihood of promotion are appropriately controlled for. The key assumption is that all group members defined as similarly situated are on average equally likely to possess the same omitted variable values. That assumption may not hold with age. An individual who has a high degree of competency (omitted variable) may be promoted more rapidly, and thus seem to be favored for being younger, than the less talented who ascend the corporate ladder more slowly. A curvilinear relationship between age and promotability may exist.
To the extent that these long-tenure employees are also not performing as well as the younger rising stars that are passing through the same units on their way to better things, there can be impact if the organization decides that it must shed its less productive people.
This view can be taken in contrast to theories involving stereotyping and implicit bias (mentioned below). Whether performance assessment is accurate or ratings are a “tainted variable” reflecting bias against older workers can make all the difference.
The Karlo court upheld decertification of the collective action and addressed three other matters regarding the testimony of plaintiffs’ experts.
Statistical correction. Campion had examined three age subgroups in addition to the 50 and above subgroup to show that RIFs increased with increased age. Was this a “hypothesis driven” demonstration of what was in the data or “data snooping” that required adjustment (such as the Bonferroni method) to the significance level because multiple tests could capitalize on chance? The district court erred in applying a merit (persuasiveness) standard when the standard for exclusion allows for less-than-perfect analyses; persuasiveness is determined at trial. This is being remanded to the district court for further consideration, so further this writer sayeth naught.
“More reasonable” alternatives. Campion provided testimony on 20 more reasonable RIF practices, given the research literature and testimony of other witnesses in this case. This stays excluded because it is not necessary. RFOA means that the issue is whether PGW’s procedure was reasonable not if there were possibly more reasonable alternatives. This is in line with what the Supreme Court held in Smith. But as in Note 1, the degree of impact might factor into a consideration of reasonableness.
Implicit bias. Greenwald offered testimony based on his research with the Implicit Association Test and review of testimony of some PGW officials regarding implicit bias against older workers. This stays excluded. The district court held that population-wide testimony was not relevant to this specific case, and the employer’s state of mind was irrelevant in an impact case. The appellate court agreed, adding that the plaintiffs had no obligation to prove any psychological mechanism just to establish a disparity. The court further added, “That is not to say, however, that implicit-bias testimony is never admissible. Courts may, at their discretion, determine that such testimony elucidates the kind of headwind disparate-impact liability is meant to redress” (p. 53).
Implications for I-O
The Karlo court’s opinion, supplemented with the references to the Paetzold and Willborn treatise, afford practitioners with a nice review of impact issues under ADEA. The this case should, however, put us all on notice to be careful about what circuit we are working in and what discrimination theory are we trying to avoid, support, or defend against.
The court touched on two issues, violation of statistical assumptions and implicit bias that could be subjects of lengthy discussion. As professionals, we should be fostering that discussion among ourselves.
Esoteric statistical methods might be necessary due to circumstances but being a methodological pioneer can have its drawbacks. One lesson from the Paetzold and Willborn discussions is that courts may exclude stuff that juries would find difficult to understand, however elegant it is. But that should not deter being proactive in application of advanced analytic methods and workforce modeling (Grossman, Cane, & Saad, 2007).
Gutman, Koppes, and Vodanovich (2011, p. 307) noted that, “In general, case law has consistently supported RIFs that are organized and well thought out.” A description of PGW’s procedure was provided by the Karlo court: broad discretion in whom to terminate, no training on the RIF procedure, no written guidelines or policies, no impact analyses, and no documentation why any individual was RIFed. We can draw the inference.
[i] ADEA has collective actions rather than class actions because the law incorporated enforcement provisions from the Fair Labor Standards Act. Such collective actions require initially unnamed plaintiffs to opt into the suit, rather than having a class action representing all persons similarly situated unless they opt out.
[ii] Other claims were disparate treatment and retaliation. The treatment claim was tossed by the district court and subsequently dropped by plaintiffs for the appeal. The court denied summary judgment for the retaliation claims, meaning these can go to trial.
[iii] The plaintiffs were employees and so whether adverse impact applies to applicants (whether the ADEA applies differently to “employees” in contrast to “individuals”) did not arise.
[iv] But see EEOC’s regulations on RFOA (EEOC, n.d). Being reasonable can include consideration of impact.
[v] Tinkham (2010) discusses a situation where average age could decrease for nondiscriminatory reasons because accurate performance evaluations used in a RIF favored younger workers. The example here is based only on seniority.
Biddle, D. A. (2011). Adverse impact and test validation: A practitioner's guide to valid and defensible employment testing, 3rd ed. West Conshohocken, PA: Infinity.
Dunleavy, E. M., & Morris, S. B. (2017). An introduction to adverse impact measurement in the EEO context. In E.M. Dunleavy & S. B. Morris, Adverse impact analysis: Understanding data, statistics, and risk. New York, NY: Routledge.
EEOC (n.d). Questions and answers on EEOC final rule on disparate impact and "reasonable factors other than age" under the Age Discrimination in Employment Act of 1967. Retrieved from https://www.eeoc.gov/laws/regulations/adea_rfoa_qa_final_rule.cfm.
Grossman, P., Cane, P. W. Jr., & Saad, A. (2007). “Lies, damned lies, and statistics”: How the Peter Principle warps statistical analysis of age discrimination claims. Labor Law, 22, 251-270.
Gutman, A., Koppes, L. L., & Vodanovich, S. J. (2011). EEO law and personnel practices, 3rd ed. New York, NY: Routledge.
Paetzold, R. L. & Willborn, S. L. (2016). The statistics of discrimination: Using statistical evidence in discrimination cases. Eagan, MN: Thomson Reuters.
Siskin, B. & Schmidt, N. (2017). Proper methods for statistical analysis of promotions. In E. M. Dunleavy & S. B. Morris (Eds.), Adverse impact analysis: Understanding data, statistics, and risk. New York, NY: Routledge.
Tinkham, T. (2010). The uses and misuses of statistical proof in age discrimination claims. Hofstra Labor and Employment Law Journal, 27, 357-394.
Barnes v. GenCorp, 896 F.2d 1457 (6th Cir. 1990)
EEOC v. McDonnell Douglas, 191 F.3d 948 (8th Cir. 1999)
Connecticut v. Teal, 457 U.S. 440 (1982)
Gross v. FBL Financial Services, Inc., 557 U.S. 167 (2009)
Karlo v. Pittsburgh Glass Works, No. 15-3435 (3rd Cir. January 10, 2017)
Lowe v. Commack Union Free School District, 886 F.2d 1364 (2nd Cir. 1997)
Meacham v. Knolls Atomic Power Laboratory, 128 S. Ct. 2935 (2008)
O’Connor v. Consolidated Coin Caterers Corp., 517 U.S. 308 (1996)
Rabin v. PriceWaterhouseCoopers, No. 16-cv-02276 (N.D. California October 11, 2016)
Smith v. City of Jackson, 544 U.S. 228 (2005)
Villarreal v. R.J. Reynolds Co. 806 F.3d 1288 (11th Cir. 2015)