Eleven American Indian plaintiffs appeal the district court’s judgment in favor of the defendants, the Governor and Secretary of State of Montana. The plaintiffs contend that the 1992 redistricting plan for Montana’s State House of Representatives and Senate dilutes the voting strength of American Indians in violation of § 2 of the Voting Rights Act of 1965, 42 U.S.C. § 1973. Plaintiffs further urge that Montana’s Districting and Apportionment Commission adopted the 1992 redistricting plan with a discriminatory purpose, an action that would also violate § 2 of the Voting Rights Act.
Following a bench trial, the district court rejected both of plaintiffs’ claims. We agree with the district court that there was Insufficient evidence of discriminatory purpose on the part of Commission members in adopting the plan, and affirm the court’s ruling on that claim. We conclude, however, that the district court’s finding that the 1992 redistricting plan did not dilute the voting strength of American Indians was based upon two errors that require its reversal. First, the district court erred in relying in part on the electoral success of Indian candidates in majority-Indian House Districts when it concluded that white bloc voting in majority-white House Districts was not legally significant. See Thornburg v. Gingles, 478 U.S. 30, 50-51, 106 S.Ct. 2752, 92 L.Ed.2d 25 (1986).1 Second, the district court erred in concluding that, under Montana’s 1992 redistricting plan, American Indians were proportionally represented. Because these two errors contributed to the district court’s ultimate finding of no dilution, we reverse the district court’s judgment and remand for further proceedings. In so ruling, we do not decide the ultimate question whether, in light of the totality of the circumstances, vote dilution has occurred; that is, whether under the 1992 redistricting plan American Indians in Montana have “less opportunity than other members of the electorate to ■ participate in the political process and to elect representatives of their choice.” 42 U.S.G. § 1973(b); see Johnson v. De Grandy, 512 U.S. 997, 1013-14, 114 S.Ct. 2647, 129 L.Ed.2d 775 (1994). We leave that finding of fact to be made in the first instance by the district court, after correction of the two noted errors.
FACTUAL BACKGROUND
A. Demographics
American Indians comprise approximately 6% of the total population of the State of Montana, according to the 1990 federal census. But the Indian population is relatively young, and comprises only 4.8% of the state’s voting age population. *1118Montana’s American Indian population is also growing more quickly than the State’s population as a whole. Between 1980 and 1990, the total population of the State increased by 1.6%, but the American Indian population of the state increased by 27.9%.
Montana’s House of Representatives consists of 100 single member districts. Each member of the House serves for a two-year term. The Montana Senate consists of 50 single member districts; each Senate district is composed of two adjoining House districts. Members of the Senate are elected for four-year terms, with half of the Senate seats up for election every two years.
Under the redistricting plan adopted in 1982, Indians comprised a majority of the voting age population in 1 of the 100 House Districts. That district included the Blackfeet Reservation in Glacier County in the northwest part of the state. Indians did not represent a majority of the voting age population in any of the 50 Senate districts. By 1990, these demographics had changed with the growth of the Indian population. The 1990 census revealed that American Indians comprised a majority of the voting age population in 4 of 100 House districts and 1 of 50 Senate districts.
B. Montana’s 1992 Redistricting Plan
Since 1972, Montana’s Constitution has granted the exclusive power to adopt a redistricting plan to a five-member Dis-tricting and Apportionment Commission. Although the Montana legislature can make recommendations to the Commission, it has no direct power over the geographic composition of legislative districts. The Commission itself is reconstituted every ten years in advance of the federal census. Commission members may not be public officials, although four of the five are appointed by majority and minority leaders of each house of the state legislature. The four Commission members select the fifth member.
The federal census in 1990 revealed that population changes in Montana between 1980 and 1990 had caused some legislative districts in Montana to become malappor-tioned, and potentially violative of the one-person one-vote requirement embodied in the Fourteenth Amendment. The 1992 Districting and Apportionment Commission therefore was required to draw a new redistricting plan. None of the five Commission members selected was an American Indian.
The Commission held twelve regional public hearings beginning on April 3, 1992. Nine of these regional hearings were preceded by planning meetings that were open to the public. All of these hearings and meetings were recorded on audio tape; portions of these tapes were summarized or transcribed in the official minutes of the Commission. Statements made by Commissioners at these hearings and meetings form the basis for plaintiffs’ claim that Commission members acted with a discriminatory purpose. American Indians appearing before the Commission presented alternative districting plans. One of these plans, referred to as the “Blackfeet-Flathead Plan,” contained an alternative districting proposal for the four challenged House districts that are the subject of this appeal.
After submitting its redistricting plan to the legislature for comment, the Commission filed its statewide redistricting plan with the Secretary of State on February 24, 1993. The plan (which we, like the district court, will continue to refer to as the “1992 redistricting plan”) then became law, and the Commission dissolved.
The 1992 redistricting plan increased the number of majority-Indian House Districts (“HDs”) from four to five (i.e., HDs 5, 6, 85, 92, and 98). See Appendix I (“1994 Montana Legislative Districts” appearing in Appellee’s Supplemental Excerpts of Record [at 20]). One of these majority-Indian districts, House District 85, is included in the four districts challenged on appeal. House District 85 contains the Blackfeet Indian Reservation in Glacier County. Of the five majority-Indian districts in the state plan, HD 85 has by *1119far the largest Indian voting age population: 66% for HD 85, as opposed to 55% in HD 98, 53% in HD 6, 52% in HD 92, and 50% in HD 5.
Table 1
House Districts by Percentage Indian Voting Age Population (“VAP”)
(* indicates district lines are challenged on appeal)
Percentage Indian VAP House District
66% *HD 85
55% HD 98
53% HD 6
52% HD 92
50% HD 5
28% *HD 73
19% *HD 86
14% *HD 74
All others less than 11%
Of the remaining House Districts in the state, only three have an Indian voting age population of 11% or more: 28% in HD 73, 19% in HD 86, and 14% in HD 74. These three districts, HDs 73, 86 and 74, are the remaining three House Districts whose boundaries are challenged on appeal.
The 1992 redistricting plan contains one majority-Indian Senate District (“SD”). That district, SD 3, is located in the southeast portion of the state, and contains an Indian voting age population of 51%. SD 3 is not a subject of this appeal.
STANDARD OF REVIEW
We review for clear error the district court’s findings of fact, including its ultimate finding whether, in the totality of circumstances, vote dilution exists in violation of § 2. Thornburg v. Gingles, 478 U.S. 30, 78-79, 106 S.Ct. 2752, 92 L.Ed.2d 25 (1986); Smith v. Salt River Project Agrie. Improvement & Power Dish, 109 F.3d 586, 591 (9th Cir.1997). We retain the power, however, “ ‘to correct errors of law, including those that may infect a so-called mixed finding of law and fact, or a finding of fact that is predicated on a misunderstanding of the governing rule of law.’ ” Gingles, 478 U.S. at 79, 106 S.Ct. 2752; Salt River, 109 F.3d at 591.
VOTE DILUTION
Before addressing the errors underlying the finding of no vote dilution, we address a threshold issue arising from the scope of appeal. At trial, plaintiffs alleged vote dilution in two separate geographic areas of the state, each involving four House Districts. The district court found no vote dilution in either area. On appeal, plaintiffs challenge the district court’s finding only with regard to the first area, which includes HDs 73, 74, 85 and 86 in the northwest part of the state; plaintiffs offered the “Blackfeet-Flathead Plan” as an alternative to the 1992 redistricting plan for this area. Plaintiffs have not appealed the district court’s finding with regard to the second area, which includes HDs 91, 92, 97 and 98 in the northeast part of the state; plaintiffs offered the “Rocky Boy-Fort Belknap-Fort Peck Plan” as a substitute for this region.2
The district court analyzed the evidence of vote dilution in the eight challenged districts taken together.3 Because the geographic scope challenged on appeal is narrower than that challenged at trial, the *1120question arises whether we review the district court’s analysis of vote dilution in the eight districts challenged at trial, or review only the record evidence of vote dilution relevant to the four districts challenged on appeal.4 We conclude that we need not decide this question; under either approach, we conclude that the district court erred in two of its subsidiary findings. To reach this conclusion, we address the evidence from both points of view.
To establish a violation of § 2 of the Voting Rights Act, 42 U.S.C. § 1973, the American Indian plaintiffs must show “that the political processes leading to nomination or election in the State or political subdivision are not equally open to participation by [American Indians, a language minority group’ protected by the Act under 42 U.S.C. § 1973aa-la(e) ] in that [their] members have less opportunity than other members of the electorate to participate in the political process and to elect representatives of their choice.” 42 U.S.C. § 1973(b). As the district court properly understood, the § 2 inquiry requires a two-step process. First, plaintiffs must show three existing threshold conditions (known as the Gingles factors): (1) the population of American Indians “is sufficiently large and geographically compact to constitute a majority in a single-member district”; (2) American Indians are “politically cohesive”; and (3) the “white majority votes sufficiently as a bloc to enable it— in the absence of special circumstances, ... usually to defeat the minority’s preferred candidate.” Thornburg v. Gingles, 478 U.S. 30, 50-51, 106 S.Ct. 2752, 92 L.Ed.2d 25 (1986) (multi-member district); see Growe v. Emison, 507 U.S. 25, 40, 113 S.Ct. 1075, 122 L.Ed.2d 388 (1993) (single-member districts).
Second, if all three Gingles factors have been established, the court must decide the ultimate question of vote dilution; it must determine whether, “on the totality of circumstances,” American Indians have been denied an equal opportunity to “participate in the political process and to elect representatives of their choice.” 42 U.S.C. § 1973(b). As part of this inquiry, the court considers a non-exhaustive list of factors set out in the legislative history to the Voting Rights Act as well as cases interpreting it. See Johnson v. De Grandy, 512 U.S. 997, 1010 n. 9, 1013-14, 114 S.Ct. 2647, 129 L.Ed.2d 775 (1994). One of these factors, “proportionality,” relates the number of majority-Indian districts to the American Indians’ share of the relevant population. Id. at 1014 n. 11, 114 S.Ct. 2647.5
We turn now to the vote dilution analysis.
A. Gingles One and Two: Compactness and Political Cohesiveness
As an alternative to the legislative boundaries drawn by the State for HDs 73, *?74, 85 and 86, plaintiffs offered the Blackfeet-Flathead Plan. That plan would create a new majority-Indian House District by enlarging HD 73. See Appendix II (“House District 1 and 2 for Flathead, Lake, Glacier and Pondera Counties” appearing in Appellee’s Supplemental Excerpts of Record [at 11]). It would also create a new majority-Indian Senate District by reconfiguring the existing HD 85, already a majority-Indian district, and combining it with enlarged HD 73.6
Under the BlaekfeetAFlathead Plan, the Indian voting age population of HD 73 would increase from 28% under the 1992 redistricting plan to 53% under the proposed plan. The Indian voting age population in HD 85 would decline from 66% to 57%. American Indians would represent 55% of the voting age population in the proposed combined Senate District.
The district court found that the population of American Indians residing on the-Blackfeet and Flathead Indian Reservations is sufficiently large and geographically compact to constitute a majority in an additional single member House District and an additional single member Senate District. The district court also found that the proposed House and Senate Districts in the Blackfeet-Flathead Plan are reasonably compact and regular, and that their shape is not materially different from that of other districts adopted by the 1992 Commission. The parties do not dispute these findings.
The State does claim, however, that the majority-Indian districts proposed by plaintiffs would be insufficient to confer effective voting power. The evidence proffered by the State does not support its claim, however, even were we to conclude that Gingles required such a finding. See Gingles, 478 U.S. at 50, 106 S.Ct. 2752 (requiring only a “majority in a single-member district.”). Badillo v. City of Stockton, 956 F.2d 884, 890-91 (9th Cir.1992), on which the State relies, does not address this factor. The district court’s finding was not clearly erroneous. The first factor of the Gingles analysis is therefore satisfied. See Gingles, 478 U.S. at 50, 106 S.Ct. 2752.
The second Gingles factor is also satisfied. American Indian voters are “politically cohesive” if they have “expressed clear political preferences that are distinct from those of the majority.” Gomez v. City of Watsonville, 863 F.2d 1407, 1415 (9th Cir.1988); see also Gingles, 478 U.S. at 51, 106 S.Ct. 2752. The State’s expert, Dr. Jeffrey Zax, presented evidence that showed American Indians were politically cohesive in more than 70% of the general elections, retention elections and ballot issue elections that he examined in the eight House Districts. Plaintiffs’ expert, Dr. Joe Floyd, presented evidence that American Indians were politically cohesive in more than 80% of the elections that he examined in which a white candidate opposed an American Indian candidate for a position in State, federal or county government.7 The district court had no need to inquire more deeply into these statistical findings. The parties stipulated that American Indian voters were politically cohesive in all eight districts challenged at trial. The parties do not challenge this stipulation as it applies to the four districts challenged on appeal.
B. Gingles Three: White Bloc Voting
Under the third Gingles factor, the court must determine whether the “white *1122majority votes sufficiently as a bloc to enable it — in the absence of special circumstances, ... usually to defeat the minority’s preferred candidate.” Gingles, 478 U.S. at 51, 106 S.Ct. 2752. In such a case, white bloc voting is said to be “legally significant.” Id. at 55-56, 106 S.Ct. 2752. The district court followed a two-step process for ascertaining the existence of white bloc voting in a particular contest: the process requires the court (1) to determine the candidate preferred by Indian voters; and (2) to determine whether whites voted as a bloc to defeat the Indian-preferred candidate. We previously employed this approach in Ruiz v. City of Santa Maria, 160 F.3d 543, 550-54 (9th Cir.1998), cert, denied, 527 U.S. 1022, 119 S.Ct. 2367, 144 L.Ed.2d 771 (1999), and Valladolid v. City of National City, 976 F.2d 1293, 1296-97 (9th Cir.1992).
The district court applied this two-step process and found that white majority voters did not usually (i.e., more than half of the time) vote to defeat the preferred candidate of Indian voters. But this finding rests on an error of method. In calculating the incidence of white bloc voting, the district court drew no distinction between jurisdictions in which Indian voters constitute a majority of the voting age population, and those jurisdictions in which white voters were in the majority. The third Gingles factor directs the court’s inquiry to those jurisdictions where there is a “white majority.” Gingles, 478 U.S. at 50, 106 S.Ct. 2752. This issue did not arise in Gingles, of course, because there were no majority black jurisdictions in the challenged plan. Johnson v. De Grandy, decided eight years later, is more instructive. There the Court faced a post-Gingles phenomenon, a § 2 challenge to a plan that already included some majority-minority districts. The Court described this kind of challenge as a claim of “partial submergence; not the chance for some electoral success in place of none, but the chance for more success in place of some.” De Grandy, 512 U.S. at 1012-13, 114 S.Ct. 2647.
In De Grandy, the Court ratified the district court’s conclusion that the third Gingles factor had been satisfied where there was “a tendency of non-Hispanic whites to vote as a bloc to bar minority groups from electing their chosen candidates except in a distnct where a given minority makes up 'a voting majority.” Id. at 1003-04, 114 S.Ct. 2647 (emphasis added) (citing De Grandy v. Wetherell, 815 F.Supp. 1550, 1572 (N.D.Fla.1992), aff'd in part and rev’d in part sub nom. Johnson v. De Grandy, 512 U.S. 997, 114 S.Ct. 2647, 129 L.Ed.2d 775 (1994)); see also id. at 1009, 114 S.Ct. 2647. Following De Grandy, we consider Indian electoral success in majority-Indian districts only in the inquiry into the totality of the circumstances — particularly proportionality and proportional representation. See id. at 1012-15, 114 S.Ct. 2647. To do otherwise would permit white bloc voting in a majors ty-white district to be washed clean by electoral success in neighboring majority-Indian districts. Such an approach would be antithetical to the directive in Gingles that legally significant white bloc voting be determined on a fact-specific “district to district” basis. Gingles, 478 U.S. at 56, 106 S.Ct. 2752.
The district court also considered white bloc voting to be probative only if the white-preferred candidate won with a minimum of 60% of the white vote. Plaintiffs contest this standard, and argue that white bloc voting is per se probative if it is sufficient to defeat the Indian-preferred candidate. See id. (relevant white bloc voting varies according to a number of factors, including “the percentage of registered voters in the district who are members of the minority group”). We assume, without deciding, that the district court employed the proper threshold. Even under this more demanding standard, the third Gingles factor is satisfied.
1. Evidence of white bloc voting
At trial, the parties presented statistical evidence of racial polarization and bloc vot*?ing by white voters in 1994 and 1996, the two election years prior to discovery in which the 1992 redistricting plan w&s in effect.8 This evidence addressed racial polarization and white bloc voting in the eight districts challenged at trial, including the four districts challenged on appeal (the four districts affected by the Blackfeet-Flathead Plan). We conclude this evidence showed legally significant white bloc voting.
The district court relied upon the statistical evidence presented by the State’s expert witness, Dr. Zax. Zax analyzed precinct election returns using a standard statistical technique known as bivariate ecological regression analysis (“BERA”). The BERA analyses compared the votes a candidate received in an election with the racial composition of the electorate, and produced estimates of the voting behavior of American Indian and white voters. The analyses did not determine the voting behavior of individual voters or determine the exact behavior of a group of voters. The BERA analyses performed by Zax were independent for each district; there was no aggregation of data across districts.
Zax produced district-specific results for the eight districts challenged at trial (including the four districts challenged on appeal). Zax first determined the number of Indians and whites of voting age in the relevant election precincts. He did so by examining the 1990 census data on racial composition of the population by census block, then overlaying a precinct map on top of the census bloc maps, and apportioning the racial population data accordingly.9 When the precinct boundaries and census bloc boundaries were not coterminous, Zax assumed the population was distributed uniformly. Once the racial composition of the election precincts was determined, Zax ran the BERA analysis to estimate minority and white voting percentages by candidate and by election contest.
Zax’s BERA analysis produced district-specific results for 268 election contests in the eight districts challenged at trial (including the 133 contests in the four districts challenged on appeal). These contests included general elections, ballot initiatives and retention elections10 at the state and federal level. Zax considered an election between two candidates to be an individual “race,” and voting with respect to a “race” in a given district was a “contest.” Thus, a “race” taking place in all eight districts (e.g., the 1996 presidential election) was considered to be eight “contests.”
Zax also performed supplemental BERA analyses on twelvé primary election and fifteen general election contests occurring in the eight districts challenged at trial in which a white candidate opposed an Indian candidate. He did so by performing a new BERA analysis on data compiled by plaintiffs’ expert Floyd. Some of these contests were also included in the general elections previously analyzed by Zax.11
2. Election contests between Indian and white candidates
Elections between white and minority candidates are the most probative in determining the existence of legally signifi*1124cant white bloc voting. See Ruiz, 160 F.3d at 553-54; Nipper v. Smith, 39 F.3d 1494, 1539 (11th Cir.1994). In this case, the Indian (and Indian-preferred) candidate was usually defeated by the white majority voting as a bloc.
In 1994 and 1996, there were 10 election contests in the four challenged House Districts (HDs 73, 74, 85 and 86) in which an Indian candidate opposed a non-Indian candidate for a position in the State or federal government. See Table 2.
[[Image here]]
Of these 10 contests, there were 7 (6 in HDs 73, 74, and 86 and 1 in Lake County) in which American Indians constituted a minority of the voting age population. In each of these 7 contests, a majority of white and Indian voters preferred a different candidate (i.e., it was a “split-preference” election). In 6 of these 7 elections (86%), the Indian candidate (who was also the Indian-preferred candidate) lost. In each of those 6 losses, over 61% of white voters supported the white candidate.
In 3 of the 10 elections (two in HD 85 and one in Glacier County), American Indians constituted a majority of the voting age population. In each of those contests, whites and Indians preferred a different candidate. In each case, the American Indian (and Indian preferred) candidate won.
If all eight of the districts challenged at trial are considered, the statistics are similar. There were 17 contests between white and Indian candidates subject to BERA analysis, 6 primary and 11 general election contests. Of the 6 contests in majority-Indian jurisdictions, the Indian (and Indian-preferred) candidate won all 6. *?Of the 11 contests in majority-white jurisdictions, the Indian (and Indian-preferred) candidate was defeated in 7 contests (64%) by a white vote in excess of 60%. All 4 of the victories by an Indian candidate in a majority-white jurisdiction were represented by the victory of Democrat Bill Yellow-tail in the 1996 Democratic primary for Montana’s sole seat in the U.S. House of Representatives. Thus, the contests between white and Indian candidates suggest white bloc voting was legally significant within the meaning of the third Gingles factor.
8. Racially-polarized elections between white candidates
In 1994 and 1996 there were 51 general election contests held in the four challenged House Districts involving two white candidates. Of these 51 elections, American Indian voters and white voters preferred different candidates in 29 contests (21 in majority-white districts and 8 in majority-Indian districts). In jurisdictions with a majority-white voting age population, the candidate preferred by Indian voters lost 14 of 21 elections (67%); in 11 of those 21 elections (52%) the Indian-preferred candidate was defeated by a white bloc vote in excess of 60%. In jurisdictions with a majority-Indian voting age population, the candidate preferred by Indian voters won all 8 contests.
During the same period, there were 5 white-white election contests for the State House of Representatives or Senate for which BERA analyses were performed. We consider these contests, occurring in the challenged districts and involving the same public office subject to challenge, to be more probative than election contests for other offices. NAACP v. City of Niagara Falls, 65 F.3d 1002, 1015 n. 16 (2d Cir.1995). Of these 5, white and Indian voters preferred different candidates in 4 of the contests. See Table 8 below.
[[Image here]]
In districts with a majority-white voting age population (HDs 73, 74 and 86), the candidate preferred by Indian voters lost all 3 split-preference election contests; in each case, white bloc voting exceeded 71%. In the 1 district with a majority-Indian voting age population, the candidate preferred by Indian voters won the sole election contest.
An analysis of white-white elections in all eight of the districts challenged at trial produces similar evidence that white bloc voting usually defeats the Indian-preferred *1126candidate. Of the 100 contests analyzed, white and Indian voters preferred different candidates in 50 elections (31 in majority-white jurisdictions and 19 in majority-Indian jurisdictions). In majority-white districts, the Indian-preferred candidate was defeated 24 of 31 times (77%); more than 60% of white voters voted as a bloc to defeat the Indian-preferred candidate in 17 of those 31 election contests (55%). In majority-Indian districts, the Indian-preferred white candidate won 15 of 19 contests (79%). If only state legislative elections are considered, the Indian-prefei'red candidate was defeated in all split-preference contests, and in 5 of those 6 instances (83%) by a white bloc voting in excess of 60%. In contrast, the Indian-preferred candidate won both of the elections in a majority-Indian district. Thus, an analysis of racially-polarized white-white contests reveals legally significant white bloc voting as defined by Gingles.
The State urges us to consider, as the district court did not, the extent to which a majority of white and Indian voters agreed on candidates in the relevant electoral contests. In Askew v. City of Rome, 127 F.3d 1355, 1385 (11th Cir.1997), it was error for the district court to consider only those elections in which black and white voters preferred different candidates. In Askew, black candidates (who were also black-preferred) ran in 33 elections. Id. at 1381. In 23 elections the white majority preferred the black candidate and the black candidate won all 23 of those contests. Id. at 1381, 1385. In the 10 contests in which the white majority did not prefer the black candidate, the black candidate lost. Id. at 1385. The district court considered only the 10 split-preference contests to be probative. The Eleventh Circuit disagreed, and held that the excluded 23 contests were probative as well. Id.
Askew has no analog here. Of the 17 contests between white and Indian candidates considered by the district court, in both majority-white and majority-Indian jurisdictions, the majority of white voters preferred the white candidate 16 out of 17 times. In the sole remaining contest, the white vote was split 50-50 between the two candidates. It is true that there was more agreement between white and Indian voters when no Indian candidates were in the race. But we have no evidence to suggest that most of these contests that the State claims to evidence racial “agreement” actually touched on issues of heightened concern to the Indian community. In fact the scant evidence is to the contrary.12 In addition, none of the contests to which the State points for support are state legislative elections. Without more, we cannot conclude that the district court erred by failing to consider the extent to which white and Indian voters did not have split preferences in white-white contests.
4. Other elections
Zax also produced evidence of white bloc voting with respect to ballot initiatives and retention elections. There were 64 ballot initiatives in the four districts challenged on appeal. Of the 64 contests, Indian and white voters preferred different options in 28 contests. In majority-white districts, white voters defeated the Indian-preferred alternative in 16 of 23 contests (70%), but did so by a bloc exceeding 60% of white voters on only 4 of 23 instances (17%). There were 5 contests in majority-Indian districts, and the Indian-preferred alternative was defeated in only one of those instances.
Results were similar if all eight districts challenged at trial are considered. Of the 128 contests, 64 were split-preference elections. White voters defeated the Indian-preferred alternative in 34 of 43 contests (79%), but in only 10 of those 43 instances (23%) did white voters vote by a bloc in excess of 60%.
*?Of the 18 retention elections, white and Indian voters did not prefer different alternative in any of those contests.
5. Gingles three: analysis of white bloc voting
Considering all this evidence in the aggregate, we conclude that the white majority in the four districts challenged on appeal “votes sufficiently as a bloc to enable it ... usually to defeat the [American Indians’] preferred candidate.” Gingles, 478 U.S. at 51, 106 S.Ct. 2752. This conclusion holds even if we assume, as did the district court, that at least 60% of the white majority must vote for a candidate to constitute a white bloc. In the contests between white and Indian candidates that are most probative of white bloc voting, the Indian (and Indian-preferred) candidate was “usually” defeated by white bloc voting in majority-white jurisdictions: in 86% of the contests in the four districts challenged on appeal and in 64% of the contests in the eight districts challenged at trial. Similarly, in contests between white candidates in majority-white districts in which white and Indian voters expressed a preference for different candidates, the Indian-preferred candidate was “usually” defeated by white bloc voting: in 52% (four districts) and 55% (eight districts) of the contests analyzed. In state legislative races between white candidates, which are the most probative white-white contests, see City of Niagara Falls, 65 F.3d at 1015 n. 16, the Indian-preferred candidate lost 100% of split-preference elections in majority-white districts as a result of white bloc voting in the four districts, and 83% of contests in the eight districts challenged at trial. In no case listed above does the rate at which Indian-preferred candidates are defeated by white bloc voting fall below 50%; in the contests that are most probative of white bloc voting, the percentages are far above that threshold. This evidence is more than sufficient to establish “legally significant” white bloc voting within the meaning of the third Gingles factor. See 478 U.S. at 55, 106 S.Ct. 2752.
In contrast, Indian and Indian-preferred candidates experienced strong electoral success in majority-Indian jurisdictions. Indian candidates (all of whom were also Indian-preferred) won 100% of the 3 contests in the four challenged districts, and 100% of the 6 contests in the eight districts. Indian-preferred white candidates won all 8 (100%) of the contests in the four districts, and 15 of the 19(79%) contests in the eight districts. As we noted above, the presence of electoral success in majority-Indian jurisdictions does not lessen the significance of white bloc voting in neighboring majority-white jurisdictions.
The State does not accept our standard for determining legally significant white bloc voting. Pointing to Abrams v. Johnson, 521 U.S. 74, 117 S.Ct. 1925, 138 L.Ed.2d 285 (1997), the State urges that white bloc voting cannot satisfy the third Gingles factor when at least 22% to 38% of white voters “cross over” and vote for the minority-preferred candidate. See id. at 92-93,117 S.Ct. 1925. We reject such a bright-line test. As the Gingles Court observed, “there is no simple doctrinal test for the existence of legally significant racial bloc voting.” 478 U.S. at 58, 106 S.Ct. 2752. The reason — “the degree of racial bloc voting that is cognizable as an element of a § 2 vote dilution claim will vary according to a variety of factual circumstances.” Id. at 57-58, 106 S.Ct. 2752; see also De Grandy, 512 U.S. at 1011, 114 S.Ct. 2647. The Gingles Court itself invalidated multi-member districts in which white cross-over voting for black candidates ranged from 8% to 50% in primary elections, and from 28% to 49% in general elections. 478 U.S. at 59, 106 S.Ct. 2752.
Nor do the “factual circumstances” of Abrams suggest that such a standard is applicable here. In Abrams the district court found that black and black-preferred candidates “received significant — occasionally overwhelming — support from both black and white voters.” Abrams, 521 U.S. at 92, 117 S.Ct. 1925. White crossover voting was so significant that two black incumbents won re-election to Con*1128gress in majority-white districts. Id. at 93, 117 S.Ct. 1925. No such facts — or findings — are present here. We decline to apply the threshold urged by the State, even on a fact-specific basis.
Finally, the State offers an alternative rationale, which we reject, for the district court’s finding on white bloc voting. Losses by Indian candidates, contends the State, can fairly be ascribed to partisan politics and not race, at least where Democratic Indian candidates lose in majority Republican districts. A plurality of the Supreme Court rejected this argument in Gingles. See Gingles, 478 U.S. at 63-65, 106 S.Ct. 2752 (plurality opinion with respect to part III.C); see also Sanchez v. Colorado, 97 F.3d 1303, 1315-16 (10th Cir.1996); Collins v. City of Norfolk, 816 F.2d 932, 935 (4th Cir.1987). In any event, we need not address this question because the evidence on the record does not support the State’s argument. Indian (Indian-preferred) candidates generally received a lower percentage of white votes than did white Indian-preferred candidates in the same district. We therefore decline to attribute white bloc voting in the four challenged districts to mere partisan politics.
C. Totality of the circumstances
Having determined that the plaintiffs have satisfied the three Gingles factors, we now address the question whether the district court correctly analyzed the factors relevant to its decision that, under the “totality of the circumstances,” the 1992 redistricting plan did not impermissi-bly impair the ability of American Indians to elect their preferred representatives. See Gingles, 478 U.S. at 43-46, 106 S.Ct. 2752; Ruiz, 160 F.3d at 550. We conclude that the district court erred with regard to the factor of proportionality. Because that error may have affected its ultimate determination of no dilution, we reverse the district court’s judgment and remand for further proceedings.
In 1982 the Senate Judiciary Committee produced a list of factors (the “Senate factors”) that a court should consider in examining the totality of the circumstances.13 This list is exemplary, and not exhaustive, and “ ‘there is no requirement that any particular number of factors be proved, or that a majority of them point one way or the other.’ ” Gingles, 478 U.S. at 45, 106 S.Ct. 2752. The most important of the Senate factors are racial polarization, “the ‘extent to which voting in the elections of the state or political subdivision is racially polarized,’ ” and proportional representation, “the ‘extent to which [American Indians] have been elected to public office in the jurisdiction.’ ” Id. at 48 *?n. 15, 106 S.Ct. 2752. A third important factor is “proportionality,” the relation of the number of majority-Indian voting districts to the American Indians’ share of the relevant population. De Grandy, 512 U.S. at 1014 n. 11, 114 S.Ct. 2647. These factors are not to be applied woodenly; as the Senate Judiciary Committee noted, and the Supreme Court has reiterated, “‘the question whether the political processes are equally open depends upon a searching practical evaluation of the past and present reality,’ and on a ‘functional’ view of the political process.” Gingles, 478 U.S. at 45, 106 S.Ct. 2752 (quoting S.Rep. No. 97-417, at 30 & n. 120 (1982) (“Senate Report”), reprinted in 1982 U.S.C.C.A.N. 177, 208) (internal quotation marks and citation omitted).
. The district court found that some of the Senate factors favored a finding of a § 2 violation. There was a history of discrimination by the federal government and the State of Montana from the 1860s until as recently as 1971. American Indians have a lower socio-economic status than whites in Montana; these social and economic factors hinder the ability of American Indians in Montana to participate fully in the political process. These two findings are not contested by the State. The district court also noted that most of the elections in the eight challenged districts involving an American Indian and a white candidate were racially polarized. Finally, it noted that in at least two recent elections in Lake County, which is within the four districts challenged on appeal, there had been overt or subtle racial appeals.
The district court also found evidence that, it concluded, supported a contrary finding. First, Montana does not have unreasonably large election districts, anti-single shot provisions, or other voting practices that enhance the opportunity for discrimination against Indian voters. Second, Montana does not have a candidate slating process. Third, Montana state officials, said the court, are “generally responsive” to the needs of American Indians.14 Fourth, the policies underlying the creation of the existing district boundaries were not tenuous. Fifth, the number of American Indians elected to the state legislature was “roughly proportional” to the American Indians’ share of the voting age population in Montana.
Finally, with respect to proportionality, the district court found that the number of legislative districts in which American Indians constitute an effective majority is “roughly proportional to the American Indians’ respective share of the voting age population in Montana.” On the strength of this and all of its other findings, and considering the totality of the circumstances, the district court found no § 2 violation.
The court’s last subsidiary finding of proportionality, however, was erroneous.15 Because the Supreme Court has left open the question of whether voting age population or total population should be the measure of proportionality, we consider both proportions. See De Grandy, 512 U.S. at 1017 n. 14, 114 S.Ct. 2647. Under the 1992 redistricting plan, 5 of the 100 House Districts (5%) are majority-Indian: this proportion is 104% proportional if the 4.8% Indian voting age population of the state is considered and 83% proportional if the 6% Indian population of the state is considered. We might consider these figures to represent “rough pro*1130portionality,” as the district court found, were it not for the corresponding proportions for the State Senate and state legislature as a whole.16 The 1992 plan for the State Senate is significantly below proportionality: 1 of 50 Senate districts (2%) is majority-Indian. This proportion is 42% proportional if the 4.8% Indian voting age population of the state is considered, and only 33% proportional if the 6% Indian population is considered. Proportionality in the combined legislature is little better. There are 6 majority-Indian House or Senate Districts out of 150(4%): this proportion is 83% proportional if the 4.8% Indian voting age population of the state is considered, and 67% proportional if the 6% total Indian population of the State is considered. These figures do not permit a finding of proportionality within the meaning of De Grandy.17
We conclude that this error may have affected the district court’s ultimate finding that, in the totality of circumstances, there was no dilution of American Indian voting strength. The district court recited several factors favoring each side, indicating that the case is sufficiently close that we cannot know whether or not the district court would have found dilution if it had correctly assessed the factor of proportionality. We therefore reverse the judgment of the district court and remand the matter for further proceedings appropriate to a- new determination of the question whether, in the totality of circumstances, American Indian voting strength was diluted by the 1992 redistricting plan, in violation of § 2.
DISCRIMINATORY PURPOSE
Plaintiffs claim in the alternative that Montana’s 1992 redistricting plan was adopted with a discriminatory purpose in violation of § 2 of the Voting Rights Act, 42 U.S.C. § 1973. Consistent with Congress’ intent in amending the Voting Rights Act in 1982, we have held that § 2 can be violated not only by adopting an apportionment scheme that has the discriminatory effect of diluting the vote of minority voters protected by the Act, but by adopting an apportionment scheme with the purpose of discriminating against those voters. See Garza v. County of Los Angeles, 918 F.2d 763, 766 (9th Cir.1990). The district court found that Montana’s 1992 plan had not been adopted with a discriminatory purpose. We conclude that this finding is not clearly erroneous, and we accordingly affirm the district court’s ruling.
As the district court understood, the standard for proving discriminatory purpose under § 2 is that set out in Village of Arlington Heights v. Metropolitan Housing Development Corp., 429 U.S. 252, 264-68, 97 S.Ct. 555, 50 L.Ed.2d 450 (1977). See Askew v. City of Rome, 127 F.3d 1355, 1373 (11th Cir.1997). Plaintiffs rely on several contemporaneous comments made by Commission members in support of their contention that the district court’s findings were clearly erroneous. Although we recognize that some of the Commissioners’ comments were inflammatory, a reasonable finder of fact could conclude, as did the district court, that there was no discriminatory purpose in the adoption of the plan. Applying other Arlington Heights factors, the district court also *1131found that the Commission did not deviate from its established criteria, that the redistricting plan significantly increased majority-Indian districts, and that the plan did not burden American Indians more than whites. After reviewing the complete record before us, we conclude that these findings were not clearly erroneous.
We also reject plaintiffs’ claim that the district court misapplied the burden-shifting rule set out in Hunter v. Underwood, 471 U.S. 222, 228, 105 S.Ct. 1916, 85 L.Ed.2d 222 (1985). Underwood requires that the burden of proof be shifted to the State only if plaintiffs establish by a preponderance of the evidence that discrimination was a “ ‘substantial’ ” or “ ‘motivating’ ” factor in the decisional process. Id. at 225, 228, 105 S.Ct. 1916. The district court concluded, without clear error, that plaintiffs did not sustain their initial burden of proof. Consequently, no burden shifting was required.
We also reject the plaintiffs’ argument that the district court improperly required plaintiffs to prove “racial animus” when all that is required under § 2 is a showing of intention to accomplish the discriminatory result. See Garza, 918 F.2d at 771. The district court, however, stated and applied the correct standard; its reference to racial animus was simply a negative response to plaintiffs’ contention that the Commissioners had displayed a racial “bias.” We conclude that the district court correctly applied the relevant law in rejecting plaintiffs’ discriminatory purpose claim.
CONCLUSION
We affirm the district court’s ruling that plaintiffs have not demonstrated that Montana’s 1992 redistricting plan was adopted with a discriminatory purpose in violation of § 2 of the Voting Rights Act, 42 U.S.C. § 1973. We conclude, however, that the district court erred in its application of the law relevant to the third Gingles factor, and that this error resulted in a clearly erroneous finding that white bloc voting was not legally significant. We also conclude that the district court erred in finding proportionality between the number of legislative districts in which American Indians constituted an effective majority and the American Indian share of the voting age population of Montana. Because that error may have affected the district court's ultimate finding that, in the totality of circumstances, there was no dilution of American Indian voting strength, we reverse the judgment of the district court and remand for further appropriate proceedings.
REVERSED and REMANDED.
. For convenience, we refer to a House District with a majority American Indian voting age population as a “majority-Indian district.” Similarly, we refer to a House District with a majoritj'-w'hite voting age population as a “majority-white district.” We refer more generally to a jurisdiction with a voting age population, the majority of which is a minority within the meaning of the Voting Rights Act, as a "majority-minority district.”
. Plaintiffs originally challenged the State’s redistricting plan in a third area, the region encompassing the Crow and Northern Cheyenne Reservations in southeast Montana. Plaintiffs dismissed this basis for their claim before trial.
. The district court was careful to ensure that statistical data relating to racially polarized voting was not aggregated across districts, as the Supreme Court has noted must not be done in a § 2 case. See Thornburg v. Gingles, 478 U.S. 30, 59 n. 28, 106 S.Ct. 2752, 92 L.Ed.2d 25 (1986). That is, each BERA (regression) analysis considered as evidence in this opinion, or the district court's order, was performed only on voting data specific to the district in question. See Part B infra (discus*1120sion of BERA analyses). These analyses did not include data from other districts where voting behavior might have different explanations.
But in calculating the incidence of bloc voting by the white majority, as § 2 requires, the district court considered the incidence of white bloc voting as a unitary proportion across all eight districts challenged at trial. For example, the district court calculated the number of times that 60% or more of white voters combined to defeat the Indian-preferred candidate in the eight districts at trial as a percentage of the total number of relevant contests in the same eight districts. This percentage, of course, will vary depending on whether the ratio includes election contests from all eight districts, or only from four. The numerical value of the percentage is important because in a successful claim of vote dilution, white bloc voting must be sufficient "usually to defeat the minority's preferred candidate.” Gingles, 478 U.S. at 51, 106 S.Ct. 2752.
. Our analysis of evidence relating to the four districts challenged on appeal is aided by findings by the district court, which we discuss below, and which are not disputed by the parties.
. Proportionality is distinct from proportional representation. Proportionality relates to "the political or electoral power of minority voters,” while proportional representation refers to the electoral success of minority candidates. Id.
. In the diagram of the Blackfeet-Flathead Plan (Appendix II), HD 2 on the diagram represents the proposed reconfiguration of HD 85 in the 1992 redistricting plan; HD 1 on the diagram represents the proposed reconfiguration of HD 73. HD 1 and HD 2 on the diagram, taken together, represent the proposed new Senate District.
. Floyd’s data suffered from a methodological infirmity in that it was aggregated across challenged and non-challenged jurisdictions. See Gingles, 478 U.S. at 59 n. 28, 106 S.Ct. 2752 ("When considering several separate vote dilution claims in a single case, courts must not rely on data aggregated from all the challenged districts in concluding that racially polarized voting exists in each district.").
. "Racial polarization’’ refers to the combined effect of the second and third Gingles factors, that is, political cohesion by the American Indian minority and white bloc voting by the white majority. Ruiz, 160 F.3d at 551 (citing Gingles, 478 U.S. at 56, 106 S.Ct. 2752).
. The U.S. Census reports data by a geographic designation known as a "bloc.” This data includes racial and ethnic composition of the population. Montana maintains precinct maps, which define where a given population votes, but which do not include data on the racial and ethnic composition of the population.
. Retention elections in the eight districts challenged at trial asked voters to cast a ballot for or against retaining a particular state judge.
. We refer to these supplemental analyses by Zax only for the purpose of discussing primary election results, which Zax did not address at all in his initial analysis.
. For example, there were 32 election contests for Supreme Court Clerk, State Auditor or Secretary of State. In none of these did white and Indian voters prefer a different candidate.
. The Senate factors include:
1. the extent of any history of official discrimination in the state or political subdivision that touched the right of the members of the minority group to register, to vote, or otherwise to participate in the democratic process; 2. the extent to which voting in the elections of the state or political subdivision is racially polarized; 3. the extent to which the state or political subdivision has used unusually large election districts, majority vote requirements, anti-single shot provisions, or other voting practices or procedures that may enhance the opportunity for discrimination against the minority' group; 4. if there is a candidate slating process, whether the members of the minority group have been denied access to that process; 5. the extent to which members of the minority group in the state or political subdivision bear the effects of discrimination in such areas as education, employment and health, which hinder their ability to participate effectively in the political process; 6. whether political campaigns have been characterized by overt or subtle racial appeals; 7. the extent to which members of the minority group have been elected to public office in the jurisdiction.
Additional factors that in some cases have had probative value as part of plaintiffs' evidence to establish a violation are: [8] whether there is a significant lack of responsiveness on the part of elected officials to the particularized needs of the members of the minority group [and 9] whether the policy underlying the state or political subdivision’s use of such voting qualification, prerequisite to voting, or standard, practice or procedure is tenuous. S. Rep. No. 97-417, at 28-29 (1982), reprinted in 1982 U.S.C.C.A.N. 177, 206-07 (footnotes omitted); see also Gingles, 478 U.S. at 44-45, 106 S.Ct. 2752.
. This finding of responsiveness of elected officials may be of “limited relevance.” Sanchez, 97 F.3d at 1325.
. The district court considered the question of proportionality within the entire state, and not within a geographic subset, for example the eight districts challenged at trial or certain counties within the state. Because the parties do not object to the scope of the district court’s inquiry, we need not decide whether the entire state is the proper "frame of reference” for a proportionality finding. De Grandy, 512 U.S. at 1021 n. 18, 1022, 114 S.Ct. 2647; see also Rural West Tenn. African American-Affairs Council v. Sundquist, 209 F.3d 835, 843-44 (6th Cir.), cert. denied, - U.S. -, 121 S.Ct. 340, - L.Ed.2d - (2000).
. Even if the ratio in the State House were considered to constitute "proportionality” in a narrow sense, it would not necessarily preclude a finding of vote dilution. Proportionality is not a "safe harbor.” De Grandy, 512 U.S. at 1016-21, 114 S.Ct. 2647.
. The De Grandy Court endorsed the district court’s finding of proportionality where Hispanics represented a majority in 9 of 18(50%) of House districts located primarily in Dade County, in proportion to a 50% Hispanic voting age population in Dade County. De Grandy, 512 U.S. at 1014, 114 S.Ct. 2647. This ratio corresponds to 100% proportionality as calculated above. Similarly, the district court found that Hispanics represented a majority in 9 of 20 House districts (45%) located at least in part in Dade County, in relation to the 47% Hispanic voting age population in that area. Id, This ratio represents 96% proportionality as calculated above.