Friend or Foe?

The Old Way

Under the OFCCP's prior practice, the key (and sometimes only) factor in any pay analysis was pay grade. The OFCCP would begin with employees in the same pay grade and examine average or median compensation for, say, men and women in that grade, often requiring employers to explain the “low” pay of any woman who earned less than the median for male employees in her grade – even though by definition half of the women should make less than the median in the absence of discrimination. Alternatively, the OFCCP demanded explanations for “significant” differences in compensation, but never defined “significant,” leaving employers to guess what pay differences it might find objectionable. This overly general method of analyzing pay also limited employers' ability to statistically explain why some employees were paid more than others; for instance, if engineers and accountants shared a pay grade, they would have been lumped together by the OFCCP in a single analysis, which only examined average pay within the grade. If the average looked bad, the employer would be on the defensive.

The New Way

Under the new proposed guidelines, the OFCCP has specifically rejected (its own) pay grade-based approach and called for a move to a multiple regression approach. Regression analysis is a statistical tool that can be used to examine how pay differs between groups after taking into account differences between those two groups. For instance, consider the simple example in Table 1 (below):

Under the old approach, the OFCCP would begin by looking at the average (or median) compensation for men and women. Here, men average $50,000 while women average $30,000 — a large difference that certainly would attract the attention of anyone using a statistical method that only looks at such simple measures. However, the obvious flaw in this approach is that it allows for no differences between these men and women to be taken into account. If it were the case, for instance, that the starting salary at this company was $20,000 and then pay increased $10,000 per year of service, this simplistic averaging analysis would not take this fact into account. However, regression analysis can do just that and can tell us what difference, if any, exists between the pay of men and women when years of service are accounted for. In this case, the type of regression analysis called for by the OFCCP would provide undeniable statistical evidence that there is no difference whatsoever in pay by gender when service is taken into account. Under the new guidelines, the OFCCP would accept this result and would not waste time and employer resources investigating the superficial salary difference in the above example.

Regression Analysis

Moreover, regression analysis can tell us whether any remaining differences are large enough to be considered statistically significant. Statistical significance simply means that the difference is large enough that it is unlikely to have arisen by chance. For example, if a regression comparison of pay shows that after taking out any effect of service, men still earn $3000 per year more than women, we will know whether the difference is statistically significant, and as the new guidelines indicate, the OFCCP will be concerned only with those differences that are statistically significant. Again, this is clearly an improvement over the old methodology, because the OFCCP has now defined the types of disparities it considers to be problematic.

Of course, if all pay systems were as simple as the one described above, there would be no need for statistical analysis, as any investigator could simply be told that pay is $20,000 plus $10,000 per year of service. Pay systems in the real world are rarely this simple. But the good news is that regression analysis can handle examples far more complicated than the one set forth above. Regression analysis can provide an estimate of the difference in pay (or promotional or hiring likelihood, for that matter) after taking into account any number of characteristics that affect pay. If pay in a particular position depends on performance, seniority, and location, it is straightforward to construct a regression model that will show exactly how each of these factors affects pay.

Consequently, these new guidelines should be viewed optimistically. First, rather than dealing with the haphazard pay grade method, the guidelines call for a statistical analysis that allows for modeling of the way in which pay is actually determined. If pay is based on years of service and performance, an employer can now present an OFCCP-approved analysis showing that there is no disparity after taking into account service and performance (even if there is a disparity without taking them into account). Second, an employer can now identify in advance any problem areas and take steps designed to ensure compliance prior to any audit. Aware of exactly what statistical tool the OFCCP will apply, employers can (and in some cases must) perform the same analysis the OFCCP will and, if necessary, either determine the cause of any disparities or remedy them with precise knowledge of the measures needed to remove the disparity.

'Tainted' Variables

Two caveats to note are that the new guidelines indicate that the OFCCP will watch for potentially “tainted” factors included in models and that care should be taken to accurately model compensation. A classic example of a potentially “tainted” variable is performance ratings. For instance, if all women receive the lowest possible rating and all men the highest rating, the OFCCP will not sanction use of performance in the regression model unless the employer can show that ratings were assigned in an objective manner. Consequently, care must be exercised in identifying which characteristics to include and whether or not the results change if certain factors are excluded. Moreover, it is possible that a regression model will yield different answers to the same question (such as, “Is there a difference in pay by race?”) depending on both the variables included in the model and the form of the model. This possibility means that employers should avoid “quick and dirty” solutions such as those offered by canned software packages, which rarely capture the facts of a compensation system drawn from the real world and may well lead to an answer different than that obtained by the OFCCP.

The Supreme Court and Disparate Impact

The recent Supreme Court decision in Smith v. City of Jackson recognized disparate impact claims in age discrimination cases and was announced by media outlets under headlines such as “Court Clears Way for More Lawsuits.” Indeed, this decision does allow for age-based disparate impact cases to proceed, which means that plaintiffs need not prove intent, but can simply point to a practice that had an adverse impact on older workers. However, as with the OFCCP change, this decision should not be viewed as a harbinger of doom. The decision also made clear (as was evident from the Court's ruling against the plaintiffs in the Jackson case) that employers can defeat a claim of disparate impact by pointing to a reasonable factor other than age (RFOA) as the reason for the decision being questioned. Moreover, this reason does not have to be the only possible basis for achieving the employer's goals, only a reasonable basis. The RFOA clause in this decision means that statistics again can be used to the employer's benefit. Consider a simple example in a termination case (see Table 2, below). In this example, 38% of employees over 40 were terminated compared with just 22% of those under 40.

Under a typical disparate impact framework, these data would lead to a conclusion of an adverse impact on older workers. A statistical expert would be called in to certify that there is a statistically significant difference, and impact would be “proved.” But, under the Jackson decision, we look for a RFOA. For example, perhaps we learn in this case that the employees were in two locations, one of which was hit harder than the other. Taking location into account, we see that the termination rate for older and younger workers is identical within each location — 50% in location one and 10% in location two. The overall disparity appears only because older workers are more likely to be in location one than location two. With the identification of this RFOA — location — the disparate impact evaporates.

Again, this is obviously a simplified example and few cases are this convenient with identical termination rates once a single RFOA is found. But even where the RFOA is not so straightforward, statistical analyses such as regression are available to determine if a difference that might initially appear to be consistent with a disparate impact can be explained by other factors – thereby meeting the Court's RFOA standard. Prior to the Jackson decision, plaintiffs could and did go into court with analyses as simple as the one described above and allege disparate impact, claiming to be unconcerned about any “explanations such as differences in termination rates by location because they had identified a disparate impact. Now such an argument should fail, provided that the employer can identify a RFOA for the difference and that statistics such as regression analysis can show that the identified factor does indeed account for the difference.

While it may seem like these two changes — the new OFCCP guidelines demanding complex statistical analysis and the Supreme Court decision allowing disparate impact age lawsuits — were bad news for employers, each presents employers with an opportunity. The new OFCCP methodology moves the agency out of the statistical dark ages and allows employers to focus on an analysis that reflects the realities of the pay process and to identify, prior to any audit, exactly what the OFCCP would find. The Supreme Court decision, by explicitly stating the RFOA standard, opens the door for employers to dismiss gross statistical evidence of disparate impact by showing statistically that a RFOA accounts for any alleged disparate impact. Both of these changes seem to indicate a movement toward standards that allow employers to present statistical evidence based on the actual operation of their businesses, and this is a clear improvement that should prevent wasted time in audits and lawsuits with no statistical support.

[IMGCAP(1)]

[IMGCAP(2)]

Recent months have delivered to employers what could be seen as a nasty one-two punch. First, the Office of Federal Contract Compliance Programs (OFCCP) announced that it planned to focus its resources on “rooting out” systemic discrimination — and unveiled proposed guidelines completely altering the way it will analyze potential compensation discrimination. The new guidelines, which require employers of a certain size to use a statistical tool called multiple regression analysis, will be enforced by a team of statisticians the OFCCP has newly hired to create the ominous-sounding Division of Statistical Analysis. Second, the recent Supreme Court decision allowing disparate impact claims in age cases could be interpreted as giving the green light to additional age-bias lawsuits by removing the hurdle of proving or even alleging intent. However, these changes will not necessarily have an adverse impact on employers, and may actually be helpful.

The OFCCP

At first glance, this new era in statistical analysis could appear to be a major headache for employers when the OFCCP comes knocking on the door, particularly given its focus on announcing statistics like an obtained 31% increase in financial remedies, and the slide-rule-wielding statisticians from the new DSA waiting to use arcane tests to determine guilt. But first glances can be misleading. While the announcements of money recovery and slide rules are not good news for employers, the OFCCP's announcement does: 1) signal a move away from the arbitrary pay grade analysis used in the past; and 2) establish guidelines so that employers know the type of analysis the OFCCP expects and will credit. Put another way, the OFCCP is moving to a more realistic way of examining compensation and is telling employers in advance just what they need to do to be in compliance. That sort of knowledge, obtained outside of an audit, is invaluable.

The Old Way

Under the OFCCP's prior practice, the key (and sometimes only) factor in any pay analysis was pay grade. The OFCCP would begin with employees in the same pay grade and examine average or median compensation for, say, men and women in that grade, often requiring employers to explain the “low” pay of any woman who earned less than the median for male employees in her grade – even though by definition half of the women should make less than the median in the absence of discrimination. Alternatively, the OFCCP demanded explanations for “significant” differences in compensation, but never defined “significant,” leaving employers to guess what pay differences it might find objectionable. This overly general method of analyzing pay also limited employers' ability to statistically explain why some employees were paid more than others; for instance, if engineers and accountants shared a pay grade, they would have been lumped together by the OFCCP in a single analysis, which only examined average pay within the grade. If the average looked bad, the employer would be on the defensive.

The New Way

Under the new proposed guidelines, the OFCCP has specifically rejected (its own) pay grade-based approach and called for a move to a multiple regression approach. Regression analysis is a statistical tool that can be used to examine how pay differs between groups after taking into account differences between those two groups. For instance, consider the simple example in Table 1 (below):

Under the old approach, the OFCCP would begin by looking at the average (or median) compensation for men and women. Here, men average $50,000 while women average $30,000 — a large difference that certainly would attract the attention of anyone using a statistical method that only looks at such simple measures. However, the obvious flaw in this approach is that it allows for no differences between these men and women to be taken into account. If it were the case, for instance, that the starting salary at this company was $20,000 and then pay increased $10,000 per year of service, this simplistic averaging analysis would not take this fact into account. However, regression analysis can do just that and can tell us what difference, if any, exists between the pay of men and women when years of service are accounted for. In this case, the type of regression analysis called for by the OFCCP would provide undeniable statistical evidence that there is no difference whatsoever in pay by gender when service is taken into account. Under the new guidelines, the OFCCP would accept this result and would not waste time and employer resources investigating the superficial salary difference in the above example.

Regression Analysis

Moreover, regression analysis can tell us whether any remaining differences are large enough to be considered statistically significant. Statistical significance simply means that the difference is large enough that it is unlikely to have arisen by chance. For example, if a regression comparison of pay shows that after taking out any effect of service, men still earn $3000 per year more than women, we will know whether the difference is statistically significant, and as the new guidelines indicate, the OFCCP will be concerned only with those differences that are statistically significant. Again, this is clearly an improvement over the old methodology, because the OFCCP has now defined the types of disparities it considers to be problematic.

Of course, if all pay systems were as simple as the one described above, there would be no need for statistical analysis, as any investigator could simply be told that pay is $20,000 plus $10,000 per year of service. Pay systems in the real world are rarely this simple. But the good news is that regression analysis can handle examples far more complicated than the one set forth above. Regression analysis can provide an estimate of the difference in pay (or promotional or hiring likelihood, for that matter) after taking into account any number of characteristics that affect pay. If pay in a particular position depends on performance, seniority, and location, it is straightforward to construct a regression model that will show exactly how each of these factors affects pay.

Consequently, these new guidelines should be viewed optimistically. First, rather than dealing with the haphazard pay grade method, the guidelines call for a statistical analysis that allows for modeling of the way in which pay is actually determined. If pay is based on years of service and performance, an employer can now present an OFCCP-approved analysis showing that there is no disparity after taking into account service and performance (even if there is a disparity without taking them into account). Second, an employer can now identify in advance any problem areas and take steps designed to ensure compliance prior to any audit. Aware of exactly what statistical tool the OFCCP will apply, employers can (and in some cases must) perform the same analysis the OFCCP will and, if necessary, either determine the cause of any disparities or remedy them with precise knowledge of the measures needed to remove the disparity.

'Tainted' Variables

Two caveats to note are that the new guidelines indicate that the OFCCP will watch for potentially “tainted” factors included in models and that care should be taken to accurately model compensation. A classic example of a potentially “tainted” variable is performance ratings. For instance, if all women receive the lowest possible rating and all men the highest rating, the OFCCP will not sanction use of performance in the regression model unless the employer can show that ratings were assigned in an objective manner. Consequently, care must be exercised in identifying which characteristics to include and whether or not the results change if certain factors are excluded. Moreover, it is possible that a regression model will yield different answers to the same question (such as, “Is there a difference in pay by race?”) depending on both the variables included in the model and the form of the model. This possibility means that employers should avoid “quick and dirty” solutions such as those offered by canned software packages, which rarely capture the facts of a compensation system drawn from the real world and may well lead to an answer different than that obtained by the OFCCP.

The Supreme Court and Disparate Impact

The recent Supreme Court decision in Smith v. City of Jackson recognized disparate impact claims in age discrimination cases and was announced by media outlets under headlines such as “Court Clears Way for More Lawsuits.” Indeed, this decision does allow for age-based disparate impact cases to proceed, which means that plaintiffs need not prove intent, but can simply point to a practice that had an adverse impact on older workers. However, as with the OFCCP change, this decision should not be viewed as a harbinger of doom. The decision also made clear (as was evident from the Court's ruling against the plaintiffs in the Jackson case) that employers can defeat a claim of disparate impact by pointing to a reasonable factor other than age (RFOA) as the reason for the decision being questioned. Moreover, this reason does not have to be the only possible basis for achieving the employer's goals, only a reasonable basis. The RFOA clause in this decision means that statistics again can be used to the employer's benefit. Consider a simple example in a termination case (see Table 2, below). In this example, 38% of employees over 40 were terminated compared with just 22% of those under 40.

Under a typical disparate impact framework, these data would lead to a conclusion of an adverse impact on older workers. A statistical expert would be called in to certify that there is a statistically significant difference, and impact would be “proved.” But, under the Jackson decision, we look for a RFOA. For example, perhaps we learn in this case that the employees were in two locations, one of which was hit harder than the other. Taking location into account, we see that the termination rate for older and younger workers is identical within each location — 50% in location one and 10% in location two. The overall disparity appears only because older workers are more likely to be in location one than location two. With the identification of this RFOA — location — the disparate impact evaporates.

Again, this is obviously a simplified example and few cases are this convenient with identical termination rates once a single RFOA is found. But even where the RFOA is not so straightforward, statistical analyses such as regression are available to determine if a difference that might initially appear to be consistent with a disparate impact can be explained by other factors – thereby meeting the Court's RFOA standard. Prior to the Jackson decision, plaintiffs could and did go into court with analyses as simple as the one described above and allege disparate impact, claiming to be unconcerned about any “explanations such as differences in termination rates by location because they had identified a disparate impact. Now such an argument should fail, provided that the employer can identify a RFOA for the difference and that statistics such as regression analysis can show that the identified factor does indeed account for the difference.

While it may seem like these two changes — the new OFCCP guidelines demanding complex statistical analysis and the Supreme Court decision allowing disparate impact age lawsuits — were bad news for employers, each presents employers with an opportunity. The new OFCCP methodology moves the agency out of the statistical dark ages and allows employers to focus on an analysis that reflects the realities of the pay process and to identify, prior to any audit, exactly what the OFCCP would find. The Supreme Court decision, by explicitly stating the RFOA standard, opens the door for employers to dismiss gross statistical evidence of disparate impact by showing statistically that a RFOA accounts for any alleged disparate impact. Both of these changes seem to indicate a movement toward standards that allow employers to present statistical evidence based on the actual operation of their businesses, and this is a clear improvement that should prevent wasted time in audits and lawsuits with no statistical support.

[IMGCAP(1)]

[IMGCAP(2)]