ABA Section of Business Law
ABA Section of Business Law
Business Law Today
November/December 1998
Reducing the risk when reducing the force
It's a question of impact and treatment
By HARRIET ZELLNERZellner is president of Integral Research Inc., a New-York based company providing litigation-support and litigation-prevention services.
Need to fire people? Prepare for litigation. The fact that reductions in force almost invariably generate claims of discrimination is, by now, well known.
It is possible to form a notion of the seriousness of this litigation risk by pooling information from two data sources: the Displaced Worker Survey (conducted bi-annually by the U.S. Census Bureau) and the Equal Employment Opportunity Commission's Charge Statistics.
According to the results of the most recent Displaced Worker Survey, some 3.7 million workers were "displaced" or "downsized" — that is, laid off on account of business conditions, not poor job performance — over the 1993 to 1995 period. Over this same period, the EEOC received some 270,000 discrimination complaints, of which one third — 90,000 or so, according to our estimates — involved a termination charge. Using these data, we would estimate that 2.4 percent of all those displaced between 1993 and 1995 filed a charge of discrimination with the EEOC. Since some people file with municipal and state agencies rather than with the EEOC, the "total" charge rate over this period was probably well above 2.4 percent.
That represents a lot of litigation risk. A similar calculation for the 1995-to-1997 period should be possible soon. No one is expecting any significant changes. It's not surprising, therefore, that statistical pre-testing of reduction-in-force (RIF) lists — the early identification of patterns in the data that might, if left unchanged, form the basis of a lawsuit — is becoming increasingly common.
This type of preventive analysis was more the exception than the rule at the beginning of the decade, but the situation appears to have changed dramatically. By the Spring of 1994, the Wall Street Journal was reporting that "... a growing number of companies are using statistical analysis as a shield, before any employees have been laid off or any lawsuits have been filed. By performing the tests in advance, the companies can adjust the layoff list to make sure it doesn't include too many people in legally protected groups." And as of March of this year, an article in Work force, a popular magazine for human resource managers, advised readers that the judicious use of such techniques could "... help lower their firms' litigation risks and increase the chances they'll never have to set eyes on a jury or courtroom again."
If the use of statistical pre-testing is to live up to such high expectations, management must ensure that the appropriate tests are applied. Where statistical analysis is concerned, one size can never be expected to fit all. The first and most important distinction to be made is between testing for disparate impact and testing for disparate treatment.
When a charge of disparate impact is brought, no intention to discriminate is alleged. The focus is entirely on outcome. The plaintiff claims that some "facially" neutral policy in connection with the reduction in force — for example, the decision to retain only those with a particular kind of training or experience — has resulted in a higher termination rate of protected-class members than of the work force as a whole. In contrast, when the charge is disparate treatment, the plaintiff is alleging that the company has intentionally discriminated in its termination decisions.
While the difference between the two types of discrimination charges may appear to be less important than what is common to them, this difference is — from a statistical point of view — critical. Because of it, two different kinds of statistical evidence and two different types of statistical test are required. A test for disparate impact may not reveal the existence of disparate treatment, even if the data — properly analyzed — show serious evidence of same. Similarly, a test for disparate treatment may not indicate disparate impact, even if such impact is, in fact, severe.
To test for disparate impact, statisticians compare the percent that protected-class members represent of the RIF list to the percent they represent of the work force as a whole. If any protected class is found to be over-represented on the RIF list, application of a simple probability test — the Fisher Exact Test is the one most commonly used — will reveal whether the degree of over-representation is "statistically significant" under the currently used evidentiary criteria.
When the charge is disparate treatment, however, the statistical situation grows more complicated. The question the statistical expert is called on to answer in these cases is no longer whether a protected class was or was not significantly over-represented in terminations. Rather, in disparate treatment cases, the court needs to decide whether or not a protected-class suffered deliberate discrimination.
How does the statistical expert test for intention? Since feelings are notoriously difficult to quantify, the statistician tests for discriminatory intention by a process of elimination. The technique used to do so is called "multiple regression." Multiple regression techniques allow us to isolate the pure (or per se) effect of gender, race or age from the effect of other factors, such as education or years with the company, that might also influence termination rates. With multiple regression, management can determine whether, in the RIF as planned, protected-class membership — all else constant — appears to increase the likelihood of termination.
The first step in the process is to compile a list of all the objective factors on which, according to management, the termination decision was based. Was formal training an issue? How about amount and type of experience? Was the firm being re-organized so that some functions or departments or divisions were being more heavily cut than others? The next step is to obtain data on each of the identified determinants of termination for every person in the employer's work force.
The final step is to "regress" termination on protected-class membership, controlling for these factors. If the regression reveals a statistically significant relationship between membership in a protected-class and termination, we conclude that the data show evidence of disparate treatment. In other words, since all the objective factors influencing termination decisions have been eliminated (statistically) as possible explanations, such a relationship will be taken to imply intentional discrimination.
We can now see why a test for disparate impact may not detect the presence of disparate treatment. Let's look at the data on ABC Inc.'s planned reduction in force shown in Table 1. The company is seeking a smaller, but more highly educated work force. Department heads have, therefore, been instructed to terminate about 25 percent of their employees, with more of the cuts coming from the less educated portion of the work force.
The RIF list is prepared and, as shown in the first panel of Table 1, a simple check for disparate impact on older workers is performed. Clearly, the data would not support such a charge. The RIF list includes exactly the same percent — 25 percent — of both age groups. Can ABC's management assume that this implies no vulnerability to a disparate treatment charge? Not at all. Let's see what the data look like when they are broken down by education group, as shown in the second panel of the table.
As can be seen there, when we look at the two education groups separately, we find that — in each of them — older workers are to be terminated at a higher rate than younger workers. It's true that the company's interest in a more educated work force is clearly reflected in the data. ABC's work force is split equally between those with and without a college degree, yet college graduates account for much less than half of the to-be-terminated group. However, it is equally true that — whatever their education — older workers will be harder hit.
If management proceeds to a multiple regression analysis of the data, they will discover this pattern before implementing the RIF list. If the company stops its pre-RIF review after determining that the RIF as planned is free of any disparate impact, it will still remain vulnerable to a disparate treatment charge.
While it's only natural for the layperson to assume that the procedures required to test for disparate treatment — being more sophisticated in nature — would somehow automatically test for disparate impact as well, we can now see why they may not. Table 2 illustrates such a case. XYZ Inc. is also seeking a smaller and more highly educated work force. Department heads are instructed to lay off 15 percent of the college-educated employees and 30 percent of the less-educated work force. The RIF list is prepared and management checks to see that exactly the same rules have been applied to each age group. And, as can be seen in the first panel, they have been. There is no evidence whatever of disparate treatment.
Should XYZ 's management proceed with this RIF list without further checking? If they do, and if a disparate impact charge is brought, they can, as shown in the second panel of Table 2, expect trouble. Older workers will, in fact, be over-represented in terminations. Not because they were being treated differently, but because they are less likely than the younger workers at XYZ to hold a college degree. In this instance, equal treatment has very unequal effects.
To sum up then, comprehensive statistical monitoring of reductions in force requires the application of two types of statistical tests: one for disparate impact and one for disparate treatment. Neither is sufficient on its own. But both together should be highly effective. To expect that statistical pre-testing will mean that management will "never have to set eyes on a jury or courtroom again" is probably overly optimistic. But that the potential benefits are enormous is undeniable.
Number Number Term II. By age group alone
Joe Davis has a college degree and this means he faces a lower risk of termination at ABC than employees with no college. But -- at ABC -- some college graduates are more equal than others. Joe, who is also over 40, faces a higher risk of termination than his younger, college-educated colleagues do: 20 percent vs. 15 percent. He will -- if "riffed" -- have grounds for a disparate treatment (but not for a disparate impact) charge.
Number Number Term Under 40
40
& over
Mona Davis has a college degree too, and -- at XYZ -- a college degree means a lower risk of termination: 15 percent vs. 30 percent. But Mona is also over 40 and older workers face a higher risk of termination at XYZ than younger ones: 30 percent vs. 20 percent. If she is among the "riffed," she will have grounds for a disparate impact (not for a disparate treatment) charge.
employed
on
RIF list
rate
I. By age and education group
Age
12,000
3,000
25 percent
Under 40
6,000
1,500
25 percent
40 & over
6,000
1,500
25 percent
College
Under 40
2,000
300
15 percent
40 & over
4,000
800
20 percent
No
college
Under 40
4,000
1,200
30 percent
40 & over
2,000
700
35 percent
employed
on
RIF list
rate
I. By age and
education group
College
4,000
600
15 percent
No college
2,000
600
30 percent
College
2,000
600
15 percent
No college
4,000
1,200
30 percent
II. By age group
alone
Age
12,000
3,000
25 percent
Under 40
6,000
1,200
20 percent
40 & over
6,000
1,800
30 percent
