Reversed by published opinion. Senior Judge CHAPMAN wrote the opinion of the court, in which Chief Judge WILKINSON and Judges RUSSELL, WIDENER, HAMILTON and NIEMEYER joined. Chief Judge WILKINSON wrote a concurring opinion, in which Judges RUSSELL and WIDENER joined. Judge WILKINS wrote a concurring opinion, in which Judge WILLIAMS joined. Judge LUTTIG wrote a concurring opinion. Judge MICHAEL wrote a dissenting opinion, in which Judges ERVIN, HALL, MURNAGHAN, and MOTZ joined.
OPINION
CHAPMAN, Senior Circuit Judge:Plaintiffs-appellants are five male professors at Virginia Commonwealth University (“VCU”) who brought this action under Title VII of the Civil Rights Act of 1964 in the Eastern District of Virginia. The appellants objected to pay raises that VCU gave to its female faculty in response to a salary equity study conducted at the university. Both parties moved for summary judgment. The district court, finding a statistically demonstrated disparity between female and male faculty salaries, denied the appellants’ motion and granted VCU’s motion. Because we find that there remains a genuine issue as to material fact, we reverse the district court’s grant of summary judgment.
I.
VCU is a state institution of higher learning located in Richmond, Virginia. In the spring of 1988, several groups called on VCU to evaluate its pay structure to determine whether female professors were victims of sex based discrimination in pay. VCU appointed a Salary Equity Study Committee to investigate, and a Salary Equity Advisory Committee to review the findings of the Study Committee.
The Study Committee chose to employ a multiple regression analysis, which compares many characteristics within a particular set of data and enables the determination of how one set of factors is related to another, single factor. The VCU study controlled for such differences as doctoral degree, academic rank, tenure status, number of years of VCU experience, and number of years of prior academic experience. Any difference in salary after controlling for these factors was attributed to sex. The study included only tenured or tenure-eligible instructional faculty at the rank of assistant professor or higher. The first regression study in the summer of 1989 showed a $1,354 difference in salaries not attributable to permissible factors. A second analysis run in the summer of 1991 showed a difference of $1,982.
Until the study, the compensation system at VCU had been based on merit alone. A professor was awarded a pay increase after a detailed annual review, if funds were available. Merit factors considered in the annual *675review were teaching load, teaching quality, quantity and quality of publications, quantity and quality of research, and service to the community (the “performance factors”). The department chair recommended pay raises to the dean, and the dean awarded raises, subject to approval from VCU’s Board of Visitors. Salaries vary widely from department to department.
The multiple regression analysis did not include the performance factors because VCU contended that these would be too difficult to quantify. VCU maintained that indirect performance variables were already included in the study in the form of academic rank, status, and experience. The study also did not take into account a faculty member’s prior service as an administrator. Administrators are paid higher wages, and faculty members retain this increase in salary when they return to teaching, thus inflating faculty salaries. Most of the faculty that had previously served as administrators were men. Furthermore, the study did not include career interruptions when measuring academic experience. Finally, the Study Committee worked under the assumption that there was no reason to suspect that female faculty members were less productive on the average than male faculty members.
After the study was completed, VCU approved more than $440,000 to increase female faculty salaries. These funds were outside of the normal salary process. The pay increases were implemented by the Salary Equity Implementation Committee made up of three women. Female faculty members had to apply for a pay increase by submitting a curriculum vitae or a narrative statement and a vitae. Of the 201 women eligible for salary review, 172 requested it. All women who requested a review received an increase in salary.
After the district court’s denial of the plaintiffs’ motion for summary judgment and the grant of VCU’s motion, the plaintiffs filed a motion to alter or amend the judgment. In this motion, the plaintiffs offered the affidavit of expert witness Dr. Fred McChesney. McChesney contended that the performance factors VCU claimed it could not quantify had in fact been included in several studies of various faculty systems, and that the inclusion of the performance factors and other variables was necessary to ensure accurate statistical data. McChesney also contended that there was data to dispute VCU’s assumption that women were as equally productive as men. In response, VCU’s expert witness, Dr. Rebecca Klemm, stated that she ran several various statistical studies with VCU’s raw data and found a salary gap to be consistent with that found in the study. McChesney never conducted a pay study himself. The district court denied the motion to alter or amend the judgment.
II.
This case comes to us on the district court’s grant of a motion for summary judgment. A motion for summary judgment should be granted only where there is no dispute as to material fact and the moving party is entitled to judgment as a matter of law. Shaw v. Stroud, 13 F.3d 791, 798 (4th Cir.), cert. denied, — U.S.-, 115 S.Ct. 67, 130 L.Ed.2d 24 (1994). We review the district court’s grant of summary judgment de novo. Tuck v. Henkel Corp., 973 F.2d 371, 374 (4th Cir.1992), cert. denied, 507 U.S. 918, 113 S.Ct. 1276, 122 L.Ed.2d 671 (1993). The evidence must be viewed in the light most favorable to the non-moving party. Ross v. Communications Satellite Corp., 759 F.2d 355, 364 (4th Cir.1985).
Under Title VII of the Civil Rights Act, employers are prohibited from discriminating on the basis of sex with respect to compensation. 42 U.S.C. § 2000e-2(a)(l) (1992). On its face, an affirmative action plan that provides for pay raises to only female faculty members violates this provision. However, courts may remedy a violation of § 2000e-2(a)(l) by instituting an appropriate affirmative action plan. 42 U.S.C § 2000e-5(g)(l). Also, the Supreme Court has determined that in certain circumstances, an employer may voluntarily establish an affirmative action plan without violating Title VII. United Steelworkers v. Weber, 443 U.S. 193, 197, 99 S.Ct. 2721, 2724, 61 L.Ed.2d 480 (1979).
*676According to Weber, an employer’s voluntary affirmative action plan is not a violation of Title VII if (1) its purpose is similar to that of Title VII, namely to “break down old patterns” of discrimination; (2) the plan does not “unnecessarily trammel” the rights of those outside the group that it is designed to protect; and (3) it is designed to eliminate a manifest racial or sexual imbalance. Id. at 208, 99 S.Ct. at 2729-30; Johnson v. Transportation Agency, 480 U.S. 616, 628-31, 107 S.Ct. 1442, 1450-52, 94 L.Ed.2d 615 (1987). The burden of establishing that an employer’s voluntary affirmative action plan violates Title VII is on the plaintiff:
Once a plaintiff establishes a prima facie case that race or sex has been taken into account in an employer’s employment decision, the burden shifts to the employer to articulate a nondiscriminatory rationale for its decision. The existence of an affirmative action plan provides such a rationale. If such a plan is articulated as the basis for the employer’s decision, the burden shifts to the plaintiff to prove that the employer’s justification is pretextual and the plan is invalid.
Johnson, 480 U.S. at 626, 107 S.Ct. at 1448-49.
The appellants assert that there is a material question of fact as to whether all the Weber requirements are met. Specifically, whether there was a manifest imbalance in pay between male and female faculty members, and whether the plan instituted by VCU unnecessarily trammelled the rights of male faculty. We agree that there is a material question of fact as to whether there was a manifest imbalance in compensation between the male and female faculty and therefore do not reach the question of whether the plan unnecessarily trammelled the rights of the male faculty.
VCU relied solely on its multiple regression analysis to determine manifest imbalance and to grant the pay increases to the female faculty. The district court also relied on the multiple regression analysis to conclude that VCU had statistically shown an imbalance between male and female faculty salaries. Therefore, the validity of VCU’s affirmative action plan depends upon the accuracy of the multiple regression analysis.
The Supreme Court dealt with the use of multiple regression analyses in Bazemore v. Friday, 478 U.S. 385, 106 S.Ct. 3000, 92 L.Ed.2d 315 (1986). In Bazemore, the Fourth Circuit disallowed the use of a multiple regression analysis at trial because the analysis failed to include all measurable factors that could have an effect on the result of the analysis. The Supreme Court reversed, holding that the multiple regression analysis should have been admitted:
While the omission of variables from a regression analysis may render the analysis less probative than it otherwise might be, it can hardly be said, absent some other infirmity, that an analysis which accounts for the major factors “must be considered unacceptable as evidence of discrimination.” Normally, failure to include variables will affect the analysis’ probativeness, not its admissibility.
Importantly, it is clear that a regression analysis that includes less than “all measurable variables” may serve to prove a plaintiffs case.
Id. at 400, 106 S.Ct. at 3008-09 (citation omitted). Bazemore was not decided at the summary judgment stage, however. The Supreme Court held that the regression analysis was to be considered in light of all the other evidence in the record. Id. at 401, 106 S.Ct. at 3009. VCU cannot rely on Bazemore at summary judgment to establish as a matter of law that the multiple regression analysis was sufficient to determine manifest imbalance.
The appellants question several aspects of the multiple regression analysis. Most importantly is the study’s failure to account for the performance factors. Bazemore and common sense require that any multiple regression analysis used to determine pay disparity must include all the major factors on which pay is determined. The very factors (performance, productivity, and merit) that VCU admittedly considered in determining prior pay increases were left out of the study. VCU maintains that it included what it called “crude proxies” for these factors, because, due to their subjective nature, the *677performance factors are not suitable for statistical analysis.1 The appellants’ expert stated that the study was not valid without adding the performance factors, and that studies performed by disinterested outside researchers “have regularly included productivity measures such- as teaching loads and publications,” and these studies have shown that productivity has a positive effect on the level of faculty compensation. We find that the questions of whether these proxies for performance were sufficient to account for merit and whether the actual performance factors could and should have been included in the study are questions of material fact.
Also at issue is the fact that the study included male faculty members who had returned from higher paying positions in the VCU administration but kept the higher salary. The study did not account for this salary differential.2 This, according to the appellants, leads to an illogical comparison involving an inflated pool of faculty members; eighty-five percent of the faculty whose salaries were increased because of prior service as administrators were male. An inflated pool can undermine the validity of a statistical study to determine imbalances. Johnson, 480 U.S. at 636, 107 S.Ct. at 1454; see E.E.O.C. v. Sears, Roebuck & Co., 839 F.2d 302, 322-24 (7th Cir.1988). Appellants’ expert, Dr. McChesney, stated that failure to include a faculty member’s status as a former administrator could easily have caused a salary differential that was not attributable to sex. Dr. Henry, the man who designed VCU’s regression study, stated that inclusion of this factor in the study would have had an effect on the study, and that if he had had the information, he would have included it. The appellants clearly produced evidence to support a finding that the pool was inflated, and this would skew the result.
The study also failed to measure the amount of time actually spent in teaching as opposed to the lapse of time since the professor began teaching. Once again, the effect of these issues on the validity of the study is a question of material fact.
Given the number of important variables omitted from the multiple regression analysis, and the evidence presented by the appellants that these variables are crucial, a dispute of material fact remains as to the validity of the study to establish manifest imbalance. Therefore, the decision of the district court granting VCU’s motion for summary judgment is
REVERSED.
. Many of the proxies for performance, such as the grant of tenure and being named departmental head, are one-time occurrences. Professors at VCU, however, continue to receive pay raises based on merit for many years after the grant of tenure or serving as a department head, so these "proxies” are not permanent, conclusive evidence of merit or performance. Merit and performance are not static — but continuing.
. The study included 770 full-time, tenure-track faculty members. Of this group, 187 were female. There were 82 members of the study poo! who were paid more because they had held administrative posts. Of this group, 71 were male.