Opinion by Judge KOZINSKI.

These cases return to us following our remand in United States v. Redondo-Lemos, 955 F.2d 1296 (9th Cir.1992) (Redondo-Lemos I). In that appeal, we reviewed a district judge’s finding that the Office of the United States Attorney for the District of Arizona was committing equal protection violations by treating male drug carriers more harshly in plea bargaining than similarly situated females. We noted that “[a] district judge who perceives a pattern of invidious enforcement has ample authority under the court’s supervisory powers to raise the matter sua sponte.” Id. at 1298 ; see also id. at 1301. We also held that, had invidious discrimination been proven, the district judge’s chosen remedy — giving defendants a sentence below the statutory minimum — would have been warranted.

On the record presented in Redondo-Le-mos I, however, we reversed, as there had been no showing of intentional discrimination; the district judge’s own observation of disparate impact “at best, established] a pri-ma facie case.” Id. at 1301 . Without more, however, it provided “an insufficient basis” for finding “that the prosecutor was motivated by a discriminatory purpose in the very case before [the court].” Id.

On remand, the district court conducted an evidentiary hearing that consisted largely of testimony by Assistant United States Attorneys, each of whom explained the reasons for various prosecutorial decisions. Defendants introduced no evidence. The court nonetheless found that the U.S. Attorney’s Office for the District of Arizona had engaged in intentional discrimination on the basis of gender. ER 21a. The court therefore again sentenced defendants below the statutory míni-mums for the offenses of which they were convicted. The government appeals.

At the evidentiary hearing, several AUSAs testified about their plea bargaining decisions, both as to these three defendants and as to the ten others the district judge cited as supporting his finding of gender-based selective prosecution in United States v. Redondo-Lemos, 754 F.Supp. 1401 (D.Ariz.1990).

The AUSA handling the Redondo-Lemos prosecution offered, in exchange for a guilty plea, to recommend the lowest sentence allowed by the Guidelines and still within the mandatory minimum. She testified that she assessed the strength of her case, followed the factors set out in the Memorandum from Dick Thornburgh, Attorney General, to All Justice Department Litigators (June 8,1989), was not motivated by defendant’s gender and knew of no office policy of plea bargaining on the basis of gender. RT of 10/19/92 at 18-20. The government worked out a tentative plea agreement with Alcaraz, conditioned on his successful completion of a polygraph examination. When Alcaraz-Peralta was unable to complete the polygraph on two separate occasions, the plea offer was withdrawn. Id. at 5. In Nolasco-Cota’s case, the government was prepared to charge him with a lesser offense if he could provide significant information, but his counsel said that Nolasco had no information to provide. Because the government considered its case against Nolasco very strong, it didn’t offer him a plea bargain. Id. at 9-10.

In the ten other cases which concerned the district judge, the responsible AUSAs testified that they based their plea bargaining decisions on the strength of the evidence, the legality of the stops and searches, the defendant’s cooperation, the level of the defendant’s involvement, and special circumstances surrounding particular defendants. See Gov’t Opening Br. at 8-12. In one case, for example, the AUSA allowed a woman who was overdue in her pregnancy to plead guilty to a misdemeanor. Because the Marshal’s Service had indicated it couldn’t give her proper medical care, she was sentenced to time served, while her male partner, who had a record, pleaded guilty and was sentenced to 90 months in prison. The AUSA explained that she would have responded the same way had the Marshal’s Service expressed concern about medical care for a male defendant. In another case, the AUSA accepted a defendant’s offer to plead guilty to a lesser offense in exchange for letting his wife, a co-defendant, go free to care for the couple’s three children. RT of 10/19/92 at 80-82.

The district court rejected these explanations as “mere general assertions that [the AUSAs] did not discriminate.” ER 16a (quoting Batson v. Kentucky, 476 U.S. 79, 94 , 106 S.Ct. 1712, 1721 , 90 L.Ed.2d 69 (1986)). .As such, the district court held, they were not sufficient to rebut the prima facie case of intentional discrimination that we held was established. In so ruling, the court overlooked the fact that the prima facie case does no more than shift to the party accused of discrimination the burden of articulating a legitimate, non-discriminatory explanation for its conduct. See St. Mary’s Honor Center v. Hicks, — U.S. -, -, 113 S.Ct. 2742, 2747 , 125 L.Ed.2d 407 (1993) (explaining nature of prima facie case and shifting burdens in Title VII cases); Batson, 476 U.S. at 94 n. 18, 106 S.Ct. at 1721 -22 n. 18 (relying on Title VII case law to define prima facie case of unconstitutional discrimination). By offering gender-neutral explanations for its plea bargaining decisions, the government fully carried this burden.

Defendants argue, however, that the government’s explanations couldn’t rebut the presumption of discrimination raised by the prima facie case because the district court just didn’t believe them. But, at the rebuttal stage, it doesn’t matter whether the district court believed the AUSAs were “actually motivated by the proffered reasons.” Hicks , — U.S. at -, 113 S.Ct. at 2749 (quoting Texas Dept. of Community Affairs v. Burdine, 450 U.S. 248, 254 , 101 S.Ct. 1089, 1094 , 67 L.Ed.2d 207 (1981)). The question is only whether the government introduced evidence “which, taken as true, would permit the conclusion that there was a nondiscriminatory reason for the adverse action.” Id. at -, 113 S.Ct. at 2748 (emphasis in original). The government introduced more than enough evidence of this sort. At that point the presumption of discriminatory intent raised by the prima facie case “simply drop[ped] out of the picture.” Id. at -, 113 S.Ct. at 2749 . The question remained whether the evidence — without the benefit of any presumption — supported a finding of discrimination.

Defendants also argue that, once the district court chose not to believe the government’s explanations, the evidence making up the prima facie case remained in place and was sufficient to support the court’s ultimate finding of intentional discrimination. But evidence sufficient to make out a prima facie case is not always sufficient to support the ultimate finding of intentional discrimination. To establish a prima facie case, a party must present only “enough evidence to demonstrate a reasonable inference of invidious discrimination.” Redondo-Lemos I, 955 F.2d at 1302 . Where the evidence behind the prima facie showing is strong, it may, standing alone, justify a finding of intentional discrimination. See Hicks, — U.S. at -, 113 S.Ct. at 2749 . But where the prima facie case is based on minimal evidence, it cannot. See Wallis v. J.R. Simplot Co., No. 92-36759, slip op. 5495, 5507, 1994 WL 224299 (9th Cir. May 26, 1994).

In Redondo-Lemos I, we authorized an unusual variant of the prima fade case: As an exercise of supervisory authority, the district judge could raise the matter sua sponte and the aggrieved defendant could proceed as if he (the defendant) had established a prima facie case of discrimination. A district judge, however, is not a witness; his observations in the courtroom are not proof. A judge who notices what he thinks may be a problem need not — indeed, cannot — present evidence of what he has seen and heard; he cannot be deposed or cross-examined; he cannot very well weigh his own credibility. The district judge’s observations can serve as a springboard for further inquiry, but they are not evidence.

A judge’s exercise of supervisory authority as authorized in Redondo-Lemos I thus has significant procedural implications by putting the burden on the government to proffer non-discriminatory explanations for its conduct and, in some circumstances, giving the defendant a chance to obtain limited discovery. Redondo-Lemos I, 955 F.2d at 1302 . But such a prima facie case does not carry the same substantive implications as a prima facie case based on evidence proffered by a party. See id. at 1301 (judge’s observations “at best, establish a prima facie case”). We therefore reject defendants’ argument that a prima facie case resting entirely on the district judge’s observations can carry their ultimate burden of proving intentional discrimination. See Wallis, slip op. at 5507.

The district judge based his finding of intentional discrimination on two other pieces of evidence: an AUSA’s testimony and two statistical studies tending to show that male defendants generally fare less well than females. The AUSA’s testimony concerned a case against a couple where the evidence was actually stronger against the woman. RT of 10/19/92 at 80-82. The AUSA nonetheless agreed to let the husband plead guilty so the wife could get off and take care of their children. The AUSA noted that, after 25 years of experience, he’d observed that it was “usually the Mexican males that will stand up and take [responsibility],” and he’d come to expect it. Id. at 81-84. According to the district court, this constituted a “[d]epar-ture[ ] from the normal procedural sequence,” Village of Arlington Heights v. Metropolitan Housing Dev. Corp., 429 U.S. 252, 267 , 97 S.Ct. 555, 564 , 50 L.Ed.2d 450 (1977), supporting an inference of discriminatory intent.

But the AUSA’s observation that, when faced with a choice, couples usually select the woman to care for the children reflects no policy of gender discrimination. While government endorsement or adoption of private discriminatory conduct that affects third parties can amount to a violation of equal protection, see Shelley v. Kraemer, 334 U.S. 1 , 68 S.Ct. 836 , 92 L.Ed. 1161 (1948), it cannot be so where the parties affected are also the very ones making the decision. The government’s compassionate practice of allowing one parent to stay with children who otherwise would be effectively orphaned is not rendered unconstitutional because the government allows the parents to decide which one is best suited to be the care-giver.

Nor do the statistics support a finding of intentional discrimination. There are only two situations where the Court has held statistics alone sufficient to prove invidious discrimination: Title VII multiple regression analyses and jury selection cases where the statistics present “exceptionally clear proof’ of intentional discrimination. McCleskey v. Kemp, 481 U.S. 279, 297 , 107 S.Ct. 1756, 1769-70 , 95 L.Ed.2d 262 (1987); Harris v. Pulley, 885 F.2d 1354, 1374-75 (9th Cir.1988). The district court here relied on two sets of statistics to show intentional discrimination in plea bargaining with male “mules” in the District of Arizona. One, which the district judge requested sua sponte from the local probation office, showed that male drug offenders in the District of Arizona were sentenced to an average of 36 months compared to 32 months for females, and that only 11% of males received straight probation compared to 35% of females. ER 11a. The other was a U.S. Sentencing Commission report showing that, nationwide, 61.5% of men who committed crimes subject to mandatory minimum sentences received them, while only 50% of women did.

This data, however, falls far short of what McCleskey requires in these circumstances. The Sentencing Commission figures reveal nothing in particular about plea bargaining practices, about Arizona or even about drug carriers. Nor do what the district judge himself called “random statistics,” ER 10a, from the local probation office prove intentional discrimination in plea bargaining. These raw comparisons take no account of the defendants’ level of involvement, prior criminal history, cooperation with the government, or individual circumstances. See McCleskey, 481 U.S. at 297 , 107 S.Ct. at 1769-70 . At most, they show disparate impact, but that alone cannot carry defendants’ burden to prove discriminatory intent. See id,.; Arlington Heights, 429 U.S. at 264-65 , 97 S.Ct. at 562-63 .

******

In Redondo-Lemos I, we stressed the extreme deference the courts must give to prosecutorial charging decisions. We reiterate this note of caution today. While a district judge has the authority and the responsibility to take a hard look at evidence of invidious discrimination by the U.S. Attorney’s Office, he must accord a presumption of constitutionality to prosecutorial decisions, and approach the inquiry with appropriate respect for the judgments exercised by officers of a coordinate branch of government. This is especially true where, as happened here, the government responds fully and promptly to the district judge’s concerns by presenting live testimony from those who made the prosecutorial decisions. The government amply demonstrated that it acted on the basis of sound, non-discriminatory criteria, and the defendants produced no evidence capable of carrying their burden of proof.

The case is therefore remanded to the district court with instructions that it sentence defendants to the mandatory minimums for their respective offenses.

REVERSED and REMANDED.

. The U.S. Attorney chose not to offer a statistical analysis, explaining through the testimony of Reynaldo Oaxaca, a professor of economics and an expert in statistics, that a valid statistical study (one that would identify and account for all the variables that went into plea bargaining decisions) would take more than two years to complete and cost more than $150,000. Gov't Opening Br. at 7. Reading our opinion in Redondo-Lemos I as requiring that “statistics ... be presented to enable opposing counsel to understand and analyze them and for the district court to make a meaningful judgment,” ER 15a, the district court criticized the government's failure to proffer such evidence. We did, indeed, note that the government could rebut the prima facie case of discrimination by use of statistics, .which we considered to be a less intrusive means for it to carry its burden than offering explanations as to individual cases and opening up its case files to discovery. In the end, however, the U.S. Attorney “must be given the chance to make whatever showing she deems appropriate_” 955 F.2d at 1302 (emphasis added). The U.S. Attorney's decision to be more forthcoming by giving evidence as to specific prosecutorial judgments is commendable. . In McCleskey , the defendant rested his claim of invidious discrimination on statistics that showed black defendants who killed white victims had the greatest chance of receiving the death penalty. The Court rejected this argument, because a raw comparison of the sentences of blacks and whites couldn't account for the many factors that inform the sentencing decision. 481 U.S. at 297 , 107 S.Ct. at 1769-70 . This observation applies equally to prosecutorial plea bargaining decisions. . The government did not have an opportunity to examine or rebut these statistics, and only learned of them when the district court filed its opinion. This does not comport with the full and fair evidentiary hearing we envisioned in Redon-do-Lemos I, in which "all affected parties [were to] have an opportunity to present evidence and otherwise participate.” 955 F.2d at 1301 ; cf. Fed.R.Evid. 201(e) (party adversely affected by court's decision to take judicial notice of adjudicative facts must be granted opportunity to be heard). Nonetheless, the government chose not to rest on this point and, instead, addressed the evidence on the merits. We therefore do the same.