concurring in part:
Virginia Commonwealth University has historically based its faculty salary decisions on merit and actual performance, as traditionally measured by teaching quality, course load, and scholarship. In this instance, however, VCU awarded salary increases to its female faculty members only, based exclusively upon the statistical results of a multiple regression analysis that concededly accounts for neither merit nor performance,1 nor at least two other major variables bearing directly on faculty salaries at the University — prior status as an administrator and actual years of teaching experience.2
Although the Supreme Court unanimously held in Bazemore v. Friday, 478 U.S. 385, 106 S.Ct. 3000, 92 L.Ed.2d 315 (1986), that a regression analysis need not include every conceivable variable affecting salary in order to be considered as evidence of discrimination, the Court left no question whatsoever that a regression analysis that does not control for at least major variables may be altogether inadmissible. Id. at 400, 106 S.Ct. at 3008-09; see also id. at 400 n. 10, 106 S.Ct. at 3009 n. 10 (“There may ... be some regressions so incomplete as to be inadmissible as irrelevant_”).
Given the undisputed facts that merit and performance are major, indeed the principal, determinants of salaries at VCU, and that the regression analysis upon which the University entirely relied in awarding salary increases to its female faculty members did not even attempt to factor in merit or performance (not to mention other major varia*682bles), I would hold on the authority of Baze-more that the statistical analysis proffered by VCU is wholly inadmissible as evidence of a manifest imbalance and therefore that the plaintiffs are entitled to summary judgment.3
The dissent is simply mistaken in reading Bazemore as “signalling” that the party challenging the validity of a regression analysis is required to conduct a separate regression in order to demonstrate that a particular variable omitted from the challenged regression is statistically significant and therefore “major.” See post at 687. To the extent that there is a “signal” as to what is required to prove a variable “major” in the footnote passage relied upon by the dissent (which passage, incidentally, does not even address this question), it is precisely the opposite of that perceived by the dissent. The Court’s criticism of the government respondents for failing to make any attempt — “statistical or otherwise” — to demonstrate that properly accounting for all important factors would have revealed the absence of a salary disparity, 478 U.S. at 403 n. 14, 106 S.Ct. at 3010 n. 14 (emphasis added), “signals,” if anything, that a factor may be shown to be major by means other than statistics. This being so, then certainly where, as here, the party relying upon a regression analysis expressly acknowledges that particular omitted variables (ie., merit and performance) are the principal determinants of the dependent variable (such as salary in this case), the party opposing the regression has carried its burden of demonstrating “otherwise” that the omitted variables are major.4
Despite the fact that VCU purports to quantify performance in its annual faculty salary determinations, VCU contends before us that performance factors are inherently subjective and unquantifiable, and that its regression analysis should not be invalidated merely because it failed to account for the unaccountable. VCU’s argument reveals a fundamental misunderstanding about its burden in a case such as this. Assuming that something akin to a McDonnell Douglas burden-shifting paradigm is appropriate in this context, see Johnson v. Transportation Agency, Santa Clara County, 480 U.S. 616, 626, 107 S.Ct. 1442, 1448-49, 94 L.Ed.2d 615 (1987), VCU has the burden of producing evidence that would overcome the plaintiffs’ showing of discrimination. The University fails to meet that burden when the only “evidence” it produces is a regression analysis that excludes major variables. Whether these variables were excluded by choice or because of practical impossibility is immaterial; the result in either case is that VCU has failed to meet its burden of production.5 At *683most, the University produced evidence of a salary differential between male and female faculty members that was attributable either to gender, or to merit, or to length of service, or to prior status as an administrator, or to other variables not considered in the regression, or to some combination of these omitted variables. It produced no evidence as to which of these variables actually caused the salary differential, much less evidence of the extent to which any of them may have done so.
VCU contends that its regression analysis is valid, even absent the inclusion of merit and performance variables, because it may be assumed that male and female faculty members perform equally on average. In many circumstances it might well be irrational, if not discriminatory per se, not to assume generally that men and women perform equally. But where the question to be answered is whether salary differentials are attributable to gender instead of to differences in performance, to assume equal performance is nothing less than to assume that which it is the purpose of the regression to determine. This is not a question of political correctness, as the dissent believes; it is, pure and simple, a question of logic, a question of whether the regression is science or mere legerdemain.
The assumption that any particular group of men and group of women must necessarily have performed equally on average, which the dissent uncritically accepts in the name of its political correctness, is conceptually flawed in any event. Just as individuals may perform differently because of their different individual capacities and the different degrees to which they realize those capacities (which the dissent concedes), so also may groups of individuals at any particular institution or in any particular workplace perform differently for the same reasons. There is, in other words, no more reason to assume that the group of women faculty at Virginia Commonwealth University performs identically to the group of men faculty than there is to assume that they do not. Indeed, it may well be that VCU’s women faculty significantly outperform their male counterparts on average, and that the regression analysis understated the salary differential attributed to gender, a distinct possibility the dissent does not even contemplate. But where the very question to be answered is whether a differential is due to gender discrimination or to differences in performance, the regression that indulges any of these assumptions is fatally flawed.
To refuse to indulge such assumptions is not to assume that men outperform women, contrary to the dissent’s suggestion; it is to assume nothing at all about the relative performance of the men and women at Virginia Commonwealth University. It is merely to recognize that without actually assessing the effect performance has on salary, one simply cannot know whether the salary differential identified in VCU’s “analysis” is in fact due to gender discrimination.
The dissent, extraordinarily, would even hold that merit itself is an impermissible basis upon which to determine salary, if a regression that properly controlled for merit and performance yielded a conclusion that the male faculty and the female faculty at VCU did not perform identically:
[E]ven if performance factors could measure and did in fact show differences between the productivity of men and women *684on the average, the only appropriate conclusion to be drawn is that performance factors improperly favor one sex over the other, not that one sex is actually more productive than the other.
Post at 691. To borrow the dissent’s own phrase, we should be far beyond such a point today.
Because, in my judgment, the only “evidence” of a manifest imbalance proffered by the University is so inadequate as to be inadmissible in a court of law, I would vacate and remand with instructions to the district court to enter judgment for the plaintiffs.
. Although perfunctorily arguing that tenure and academic rank are "crude proxies” for merit and performance, VCU appears to realize, as does the dissent, that tenure and academic rank are more measures of length of service at or above a minimum level of performance than of relative merit or performance above that minimum. Moreover, tenure and academic rank are not in any sense measures of either relative performance after tenure or comparative performance among individuals of the same academic rank.
. Since administrators retain their higher salaries when they return to teaching, and since most prior administrators included in the study were men (the discriminatory implications of which, if any, are not at issue in this case, contrary to the dissent's assertion, see post at 688 n. 8), the failure to control for administrative experience would necessarily increase the amount of salary differential that the regression ascribed to gender discrimination. Similarly, the failure to account for breaks in service would, to the extent of any correlation between breaks in service and gender, necessarily attribute to gender some salary differential that was in fact caused by breaks in service.
.If the district court determines on remand that tenure and academic rank are not adequate proxies for merit and performance, and that merit and performance are major variables which the regression analysis therefore excluded, it would yet be appropriate under our opinions issued today for the court to grant summary judgment for the plaintiffs. Judge Chapman holds only that it is a disputed issue of material fact whether tenure and academic rank are legitimate proxies for merit and performance; if they are not, presumably he would hold, as I would even at this juncture, that the University’s regression is inadmissible. See ante at 676-77 Judge Wilkins also fully acknowledges that an analysis is inadmissible if it fails to account for major variables; he simply views the McChesney affidavit as insufficient to establish that a major variable, as opposed to merely a measurable one, was omitted from the regression. See ante at 679 n. 1. Thus, nothing in any of the opinions in which a majority of the court concurs forecloses summary judgment for the plaintiffs on remand; indeed, if anything, our opinions reinforce the possibility of such a holding.
. The dissent also interprets the Court’s observations that the salary disparities in Bazemore were "statistically significant” as confirming that an excluded variable may only be deemed "major" if it is shown by separate regression analysis to be statistically significant. See post at 687 (citing Bazemore, 478 U.S. at 399 n. 9, 401, 404 n. 15, 106 S.Ct at 3008 n. 9, 3009, 3010 n. 15). Of course, the existence vel non of a statistically significant salary differential speaks not at all to the question of whether, in order to be considered "major," an omitted variable must be proven to have a statistically significant effect by means of a separate regression analysis, rather than by “other” means.
. Because the plaintiffs in this case offered direct, unrebutted proof of discrimination, rather than mere circumstantial evidence from which a prima facie case of discrimination may be inferred, I doubt whether the paradigm of McDonnell Douglas Corp. v. Green, 411 U.S. 792, 93 S.Ct. 1817, 36 L.Ed.2d 668 (1973), is even appropriately invoked. Even if it is, it would seem that after City of Richmond v. J.A. Croson Co., 488 U.S. 469, 109 S.Ct. 706, 102 L.Ed.2d 854 (1989), *683and Adarand Constructors, Inc. v. Pena,-U.S. -, 115 S.Ct 2097, 132 L.Ed.2d 158 (1995), there is a substantial question whether an affirmative action plan such as VCU’s can serve to meet the University’s burden of producing evidence of a nondiscriminatory rationale for its facially discriminatory actions. That is, there would seem to be a serious question whether Johnson remains good law, to the extent that the Court there allowed an affirmative action plan to satisfy an employer’s burden of production under step two of the McDonnell Douglas framework and imposed upon the party challenging such a plan the burden of proving its invalidity. See Johnson, 480 U.S. at 626-27, 107 S.Ct. at 1448-49. To accept that an affirmative action plan constitutes a nondiscriminatory rationale for discriminatory action (and thereby to require the person challenging the plan to establish its invalidity), is to accept that so-called "benign” discrimination is not discrimination about which the antidiscriminatoiy laws — be they Title VII or the Equal Protection Clause — are concerned. However, the entire premise of Croson and Ada-rand is that "benign” discrimination is discrimination nonetheless, to be subjected to the same exacting level of judicial scrutiny as "invidious" discrimination.