INTRODUCTION
Debates over White racial attitudes have been enduring, heated, and often complex. There is still no consensus over how much racial attitudes still matter in American politics. And there is even less consensus over the shape that these attitudes take and the relationship between these attitudes, policy preferences, and political behavior.
Over the last fifty years there have been immense changes in the racial attitudes of White Americans. A half century ago, White Americans, in both the North and the South, openly expressed their racist beliefs and supported policies that reflected and reinforced such prejudices. Since then, however, such expressions of White racism have declined dramatically. According to Schuman et al.,
On questions concerning principles of equal treatment of blacks and whites in the major public spheres of life (jobs, schools, residential choice, public accommodations, transportation), there has been a strong and generally steady movement of white attitudes from denial to affirmation of equality—so much so that some questions have been dropped by survey organizations because answers were approaching 100% affirmation (Schuman et al., 1997, p. 191).
Yet, for all these changes, the progress toward racial equality has been incomplete. Decades after the civil rights victories of the 1950s and 1960s, American schools and neighborhoods remain highly segregated, and Blacks still face persistent economic inequalities.2
On school segregation, see Orfield (2001). On residential segregation, see Massey and Denton (1993). For overall analyses of the persistence of racial inequality, see Smelser et al. (2001).
Others, however, question both the genuineness and significance of changes in White racial attitudes. These analysts acknowledge that the “old racism,” based upon biological notions of Black inferiority, has largely disappeared, but they claim that White attitudes regarding racial equality remain ambivalent or even hostile. For example, while most Whites express strong support, in principle, for racial equality, support is much lower for most measures designed to achieve racial equality as a fact. These analysts also suggest that, for reasons of social acceptability, many Whites mask their true racial attitudes in public opinion surveys. They argue that a “new racism,” based upon anti-Black affect and a belief that Blacks refuse to play by the same rules as other Americans, now exerts a powerful influence on Whites' attitudes toward a range of racially charged policies, including school desegregation, affirmative action, and social-welfare spending.3
Examples of this research include Kinder and Sanders (1996); Sears et al. (2000a); Kinder and Mendelberg (2000); Bobo (2000); and Sidanius et al. (2000).
A substantial challenge for this argument has been the inherent difficulty of separating anti-Black affect and policy position as explanations for survey responses. Hutchings and Valentino, in their thoughtful review of the state of recent racial attitudes research,4
For a valuable, but somewhat less current, treatment of this debate, see also Sears et al. (2000b).
Hutchings and Valentino also highlight the complexities involved in using surveys to understand racial attitudes: Attitudes expressed in public regularly differ from those expressed in private; those expressed to one interviewer may differ from those expressed to another; and expressed attitudes may change after exposure to different racial primes or cues.
The causes of these phenomena are still being explored. In one innovative experiment, Krysan (1998) addresses the first phenomenon with a set of experimental treatments designed to distinguish between two potential causes: social desirability bias (reluctance to give socially disfavored responses) and acquiescence bias (reluctance to disagree with the interviewer). What emerges strongly from this work is the importance of context: “substantive interpretations about the nature and correlates of white racial attitudes are inextricably linked with the method of measuring them, and interpretations of contemporary racial attitudes must take into account the context in which they are expressed” (Krysan 1998, p. 510).
HOW SYMBOLIC REFERENDA CAN EXPOSE RACIAL ATTITUDES
Debates over White racial attitudes have revolved largely around evidence gleaned from public opinion surveys. Interpreting these survey results has been complicated, however, by the complex effects of context. Despite the use of innovative experiments and methods to circumvent this problem, major questions remain regarding the extent and political significance of White racial attitudes. Thus, it is important to seek data from other, nonsurvey expressions of these attitudes. Furthermore, as it becomes increasingly clear that the expression of racial attitudes cannot be divorced from context, it is particularly relevant to political science that these data be embedded within the context of a political action.
Data from two recent elections offer a unique opportunity to explore White racial attitudes in a real political context. In 1998 and 2000, South Carolina and Alabama, respectively, held statewide referenda to repeal the provisions in their state constitutions banning interracial marriage. Both provisions dated back to the turn of the century (1895 for South Carolina; 1902 for Alabama), when such provisions were seen as integral to the maintenance of White racial superiority (part of a system of provisions known as the “Jim Crow” laws). Gunnar Myrdal (1944 [1964]) explained why:
The ban on interracial marriage has the highest place in the white man's rank order of social segregation and discrimination. Sexual segregation is the most pervasive form of segregation, and the concern about “race purity” is, in a sense, basic. No other way of crossing the color line is so attended by the emotion commonly associated with violating a social taboo as intermarriage and extra-marital relations between a Negro man and a white woman. No excuse for other forms of social segregation and discrimination is so potent as the one that sociable relations on an equal basis between members of the two races may possibly lead to intermarriage (Myrdal 1944 [1964], p. 606).
As a consequence, antimiscegenation clauses were quite popular, and not only in the South; in 1945, thirty states (including every southern state) had written them into their constitutions or statute books. Over the next two decades, changing racial norms led about half of these states (all outside of the South) to remove such provisions. Finally, in Loving v. Virginia (1967), the U.S. Supreme Court struck down the remaining laws as an unconstitutional violation of the equal protection clause of the Fourteenth Amendment.
With such laws now a dead letter, most states removed them from the books, and by the mid-1990s only those in South Carolina and Alabama remained. In both states, Black state legislators began efforts to remove the laws. Opponents of the antimiscegenation clauses condemned them as embarrassing and offensive relics of a bygone era, and argued that their removal would symbolize racial progress. Since the antimiscegenation provisions were part of each state's constitution, removal required approval by a majority of voters in a statewide referendum. The issue came before the voters in South Carolina in 1998, and in Alabama in 2000.
These referenda offer a useful context for exploring White racial attitudes. First, unlike survey research, referenda are conducted in the privacy of the voting booth. Consequently, social acceptability pressures are greatly reduced, since voters are free to express their true opinions without worrying how their responses might reflect upon them.
Second, support for keeping the antimiscegenation provisions is difficult to explain, absent racial bias. As mentioned previously, antimiscegenation was an important component of the Jim Crow order. And, unlike issues such as affirmative action, school busing, or social-welfare spending, support for antimiscegenation laws reflects no race-neutral sentiments, such as beliefs in limited government, individualism, or free enterprise. Furthermore, the remaining antimiscegenation measures are purely symbolic since the Supreme Court declared them unconstitutional in 1967. Removing them from the books costs nothing and imposes no additional burden on the citizens of these states.
The racial nature of these referenda can be seen in the nearly total lack of opposition to them in the media and state legislatures. In South Carolina, the House of Representatives approved deleting the ban on interracial marriages by a vote of ninety-nine to four and by a unanimous voice vote in the Senate (Gaulden 1998a, b; Rawls 1999; Johnson 2000). In Alabama, both houses passed the bill unanimously. Furthermore, in both states, few public voices were raised in support of the ban, and most major public officials and newspaper editorials called for its repeal. A search of Lexis-Nexis for news articles on the referendum campaigns shows that the only group to come out in support of the ban in either state was the 200-member Alabama branch of the Southern Party, “which reveres the Old Confederacy and wants the South to secede again” (Johnson 2000). It therefore seems clear that a person who knowingly votes against the repeal harbors some racist sentiments.
Third, these referenda offer a clear contrast to survey research responses. Around 1960, approximately 96% of Whites disapproved of racial intermarriage and 62% supported laws against it (Schuman et al., 1997, pp. 104–106). Since 1972, the General Social Survey (Davis et al., 2000) has asked respondents whether or not they support laws banning interracial marriage.5
The exact question wording was: “Do you think there should be laws against marriages between (Negroes/blacks/African Americans) and whites?” Similarly, 60% of White respondents to a Gallup poll taken the previous year reported disapproval of such marriages. In subsequent Gallup polls, White disapproval dropped to 54% in 1978, 42% in 1991, and 29% in 2000 (New York Times Correspondents 2001).
Outright approval of interracial marriage is likely to be somewhat weaker than opposition to laws banning interracial voting. For example, the recent New York Times study of How Race is Lived in America describes the confusion of an elderly White lady in a racially mixed Pentecostal congregation near Atlanta who asked, without apparent animus, whether “scripture permitted mixed marriages” (p. 15). Nevertheless, a poll conducted by the same reporters in 2000 suggested that 63% of Whites approved of interracial marriage, while 29% disapproved (New York Times Correspondents 2001).
While the General Social Survey (GSS) does not support breakdowns by state, regional data is revealing: In the South Atlantic region, of which South Carolina is a part, 86.5% of Whites (and 94.5% of non-Whites) opposed laws barring interracial marriages. In the East South Central region, of which Alabama is a part, opposition was a bit lower, at 69.8% (90.4% for non-Whites).7
Figures compiled using the General Social Survey (GSS) Cumulative File (Davis et al., 2000). Non-responses are deleted before computing these percentages. The South Atlantic region includes South Carolina, Delaware, Maryland, West Virginia, Virginia, North Carolina, Georgia, and Florida. The East South Central region includes Alabama, Kentucky, Tennessee, and Mississippi. Also note that the number of non-Blacks in each region per year is relatively small.
Figures compiled from the survey data provided to us by Keith Nichols at the University of South Alabama.
By examining the racial breakdown of the vote on these referenda, we can determine the extent to which these public opinion surveys accurately measure the racial attitudes of White respondents. In both South Carolina and Alabama, the measures passed with approximately equal levels of support, 61.9% and 59.5%, respectively. Unfortunately, no exit polls were conducted to determine how Blacks and Whites voted on the issue. Nonetheless, the voting and election data from these states allow us to estimate such a breakdown. Because these states identify the race of both registrants and voters, it is possible to determine the number of Blacks and Whites who voted in each precinct. Furthermore, Blacks and Whites make up the overwhelming majority of voters in both states, making a racial breakdown easier to determine. According to the 2000 Census, 96.7% of South Carolina's population and 97.1% of Alabama's were categorized as either Black or White.9
South Carolina data source: U.S. Census Bureau, State and County QuickFacts. Data derived from Population Estimates, 2000 Census of Population and Housing, 1990 Census of Population and Housing, Small Area Income and Poverty Estimates, County Business Patterns, 1997 Economic Census, Minority- and Women-Owned Business, Building Permits, Consolidated Federal Funds Report, 1997 Census of Governments. South Carolina data available at 〈http://quickfacts.census.gov/qfd/states/45000.html〉 (accessed October 10, 2006); Alabama data available at 〈http://quickfacts.census.gov/qfd/states/01000.html〉 (accessed October 10, 2006). The data were collected in the fall of 2002.
DATA COLLECTION
We collected data on the referenda results and the number of Blacks and Whites who voted in each precinct in South Carolina and in Alabama. For South Carolina, we obtained our data from the State Election Commission.10
This data is available at 〈http://www.state.sc.us/scsec/stats.html〉 (accessed October 10, 2006).
In addition, some precincts were unanimous—voting 100% for one option. Following King's (1997) suggestion, we ran separate estimates of every model, excluding these precincts. The results were not, however, detectably different, so we include these unanimous precincts in the results below.
It is worth noting that voting in absentee ballots did differ from voting in nonabsentee ballots. The absentee percentage of No votes as a fraction of Yes and No votes was approximately 4.2% lower in South Carolina and 9.1% lower in Alabama than for the nonabsentees. It is not possible to include the absentees in our analysis because racial and abstention variables are missing for these precincts. However, even assuming for the sake of argument the most extreme case, that absentees were 100% White and had 0% abstention, the difference would not be large enough to change our substantive conclusions.
For Alabama, the data were collected from a variety of sources. Most precinct-level election returns were obtained from the Elections Division of the Secretary of State's Office.13
This data is available from the commission's website: 〈http://www.sos.state.al.us/election/2000/2000.htm〉 (accessed October 10, 2006).
Two of the remaining counties, Chambers and Jefferson, had data only on the race of registered (as opposed to actual) voters, and the third, St. Clair, had no data available. Fortunately, St. Clair County is very small (only 20,000 referendum votes were cast there) and could be safely dropped from the analysis. Chambers is even smaller, and the number of precincts seemed too small to affect any conclusions. Jefferson had 180 valid precincts, which we used for independent-model estimations, pooling only the final results.
Inferring Individual Voting Behavior from Precinct Data
In order to compare survey behavior and voting behavior, one needs to infer the behavior of individual voters from aggregate election results. In this section, we describe the details of these inference methods, often termed ecological or cross-level methods.
The fundamental problem of cross-level inference is that relations observed in aggregate data do not necessarily correspond to relationships among individual behaviors, a problem often referred to as “the ecological fallacy.” For example, a correlation between the percentage of White population in a county and the percentage of that county voting against an amendment does not necessarily imply that individual Whites are more likely than non-Whites to vote against the amendment.
Since this problem was discovered, social scientists and voting-rights litigation have used three basic methods to deal with cross-level inference problems: bounds analysis, homogenous precinct analysis, and ecological regression. [For a detailed but not overly technical explanation of these methods and their application in social science, see Kousser (2001).]
Bounds analysis (also known as the “method of bounds”) computes the logical bounds on individual behavior implied by the percentages for each aggregated unit. For example, in a district that is 80% White, and in which only 10% of the voters cast ballots for the constitutional revision, it is numerically impossible for more than one-eighth of the Whites to have voted for the revision. For each precinct, we can calculate similar bounds for both Whites and non-Whites. Unfortunately, these bounds are, in most cases, so wide as to be substantively uninformative.
Homogenous Precinct (HP) analysis involves looking at only those precincts where the population is nearly uniform—e.g., nearly all White or nearly all Black (Grofman et al., 1992).14
There is some variation in how homogenous districts need to be in order to be considered “practically” uniform. In this study, we define “uniformity” to be at least 95% White or non-White.
Ecological regression [which is also commonly known as “Goodman's regression,” after Leo Goodman (1953)], has been widely used in voting. Goodman's regression typically involves running a regression of percentage of voting for a particular party on the percentage of Blacks across all districts. This method relies on the (somewhat implausible) assumption that the voting behavior of members of each of the groups is uniform across all precincts. In addition, Goodman's regression ignores the logical numerical bounds implied by each precinct, and thus can produce impossible (e.g., negative) estimates of turnout and voting.
Recently, King (1997) has reinvigorated this area of analysis by introducing an advance on these methods, what he calls Ecological Inference (EI), which combines ecological regression and bounds analysis.15
Just prior to King, Achen and Shively (1995) extended Goodman's regression using an explicitly modeled quadratic term. They demonstrate that this ameliorates the effects of aggregation bias. This model assumes that one behavior of either Blacks or Whites (but not both) can vary at the precinct level. We do not report results for this quadratic model. However, we did repeat our analysis using this model, and it did not yield results substantively different from those we present.
King's method has the unique advantage of producing separate estimates for each precinct, as well as overall estimates (King 1997).16
King's solution specifically applies to 2 × 2 tables of aggregate data. King recommends a multistage approach to solving the more general R × C case (King 1997, Section 8.4 and Chapter 15), which essentially involves iteratively applying the estimation to data aggregated through a previous 2 × 2 estimation. An alternative is the one-stage aggregated method described in section 8.2. We report the results of the recommended two-stage estimation. However, we ran the one-stage method as a check, and the estimates it produced were substantively the same.
In addition, our results turned out to be robust across a range of plausible covariates. See Correlates of the White Vote, below.
In particular, we examined the estimated aggregation bias, Palmquist inflation factor, return codes, resampling indicator, and log likelihood parameters for each run. We examined the posterior plots, fit plots, nonparametric plots, Goodman plots, Bias plots, and tomography (“TomogS”) plots for goodness of fit. To guard against numerical issues, we inspected the likelihood optimization trace, checked the Hessian inversion method, checked for warnings, and replicated the results using the highest precision settings (central derivatives, eighty-bit precision, and most accurate bivariate cumulative normal distribution method).
See King (1997, Chapter 16) and the EI software documentation for some recommended diagnostics. Also see Cho (1998) for cautions on model specification, and Altman et al. (2004), for a discussion of computational issues and additional diagnostics.
We tested sensitivity to 1% uniform noise on the dependent and independent variables.
FINDINGS: LARGE DIFFERENCES BETWEEN VOTING AND SURVEYS
Table 1 shows the vote against reform in each state. It is clear that Whites vote against reform at more than twice the rate of non-Whites. Since the different methods produced similar estimates for the mean aggregate propensities, we report the range of estimates produced by the other methods, but found our discussion on the estimates produced by the more sophisticated EI method.22
Note that the EI estimates we discuss are the least favorable to our overall finding that White voting behavior and stated opinions significantly differ; thus our conclusions are robust across estimation methods. Note also that the standard errors for the EI method are in general significantly smaller than the range across methods reported in Table 1.
Vote Against Reform—Estimated proportion of White voters voting against constitutional reform, using EI method, as a proportion of total White vote for and against reform. Note: Numbers in brackets show the range of estimates produced across all estimation methods.Based on EI two-stage estimates, as reported in the Appendix. The proportion reported for Alabama is a linear population-weighted combination of the two separate EI point estimates for Jefferson County and for all other Alabama counties.
These results indicate that approximately 40% of Whites and 23% of non-Whites voted No in South Carolina. In Alabama, approximately 49% of Whites and 8% of Blacks voted No.23
Roll-off and turnout were not unusual. In South Carolina, which had four amendments on the same ballot, overall roll-off ranged from 21% to 25% (weighted), with roll-off on amendment four being the least. See the Appendix for estimates of roll-off by race in each state.
This was a telephone survey of 405 adult residents of Alabama, conducted September 4–6, 2000, by USA Polling Group; margin of error is +/−5% at the 95% confidence level. Data provided to the authors by Keith Nichols of the University of South Alabama.
White support for laws banning interracial marriage, as a proportion of total White support and opposition
If voters were more likely than nonvoters to support bans on interracial marriage, then some of this difference could be explained by differences in the populations sampled. However, when we compared the opinions of likely voters in the GSS (those who had voted in the previous presidential election) to the rest of the GSS population as a whole, we found that likely voters (both overall, and in each of the relevant regions) were more likely (albeit only very slightly) to oppose laws banning interracial marriage.25
We used GSS results from 2000, and results based on GSS 1998 were also substantively equivalent. This difference was both statistically and substantively negligible. We also compared voters and nonvoters by region: the difference for South Carolina was 8.2% and the difference for Alabama was −4.2%. However, both of these regional subsamples are necessarily small, and we found (using bootstrap methods) the differences between voters and nonvoters in them not to be statistically significant either.
Distribution of the Vote Against Reform
Estimates of statewide proportions are, in this case, an oversimplification. As Figure 1 and Figure 2 show, non-White behavior is more uniform than is White behavior. In each precinct, the overwhelming majority of non-Whites voted for reform. In contrast, White behavior varied more dramatically: Whites in some precincts voted overwhelmingly against reform, while in other precincts Whites voted overwhelmingly for it.
Distribution by unweighted precinct of White and non-White propensities to vote against reform in Alabama, as a proportion of the total vote
Distribution by precinct of White and non-White propensities to vote against reform in South Carolina as a proportion of total vote
White voting was not uniform across precincts, as Figure 1 and Figure 2 show. For Alabama, we were also able to obtain sufficient information to match precincts to U.S. Census voting tabulation districts (VTDs) and thus map the geographic pattern of the White vote at the precinct level, as shown in Figure 3. In general, the map seems to show somewhat higher levels of White No votes in the traditional “Black belt” areas of the central and western part of the state.
Proportion of Whites who voted against the repeal of the antimiscegenation provision in Alabama, as a proportion of total White vote for and against reform, grouped by voter tabulation district
Correlates of the White Vote
Previous research suggests that White racism may be correlated with the level of minority concentration (the so-called “Black threat”), urbanization, education, or age.26
On the impact of age and education, see Schuman et al. (1997). On urbanization and “Black threat,” see Voss (1996, 2000) and Hutchings and Valentino (2004).
In particular, we used demographic data from the U.S. Census Decennial SF3 files to calculate, by county, the percentage of the county for each of the following categories: non-White, urban residence, Whites that had attended some college, and Whites and non-Whites sixty-five years of age and older.27
Although some relevant demographic data is available at the block-group level of aggregation, it is not generally possible to merge it cleanly with voting data, since voting precincts and block-group boundaries often overlap in complex ways. Thus, we were forced to use county-level aggregates.
To explore the relationship between White voting and these explanatory variables further, we used the method of ordinary least-squares, weighted on the standard errors of the betas, to regress the effect of the explanatory variables on the EI estimates (this technique is known as EI-W).28
Although this is, essentially, the only practical technique for estimating the relationship ecological effects in the presence of multiple covariates, some caution is warranted with this technique, as it can be biased where the precinct bounds are insufficiently narrow. Following the recommendation of Adolph et al. (2003), we apply their diagnostic (as illustrated on p. 90) to determine whether EI-W is likely to be substantially unbiased in our analysis. Fortunately, our data provides tight enough bounds on βw that both the EI-W regressions for South Carolina and for the bulk of Alabama are well within the region of estimates expected to be “approximately unbiased.” The diagnostic for Jefferson County is still within the “approximately unbiased” region, but near its edges, more caution may be warranted with respect to EI-W analysis.
Technically, the urbanization variable was statistically significant; however, we would caution against attaching any substantive meaning to this, given an estimated coefficient that was close to zero and the poor model fit.
Covariates of White voting in Alabama (using the EI-W model)
Our analysis of Alabama did suggest that Whites in more rural counties were more likely to vote No, and Whites in counties where the population is older, more Black, or more educated were less likely to vote No. The large negative coefficient on “percentage old” (i.e., of sixty-five or more years of age) is surprising, given the usual positive correlation between age and conservative racial views (Schuman et al., 1997). However, a great deal of caution is required in interpreting these results, as they are essentially driven by Jefferson County, the home of Birmingham and the largest county in the state. When Jefferson County is excluded, the model fit drops to negligible levels, the coefficients shrink by roughly two-thirds, and the significance vanishes for all but the intercept and percentage rural.30
In addition, the independent variables presented in this table must be statistically interpreted as contextual effects, e.g., the effect on an individual White voter of living in a county that has a higher proportion of minorities, in a county that has a higher proportion of rural area, or a county that contains a higher proportion of educated voters. They are not, e.g., unbiased predictors of the effect of being more educated on an individual White voter. See Greenberg et al. (1989) for cautions on the latter interpretation.
That we did not find these explanatory variables to have had strong effects on the White vote is perhaps not surprising, given the coarseness of the demographic data. Since fine-grained demographic data are not available, we can only conjecture that, if demographic information were available for each precinct, then stronger relationships would emerge.
DISCUSSION
As these results suggest, race was a somewhat more important factor in the Alabama vote than it was in South Carolina. In South Carolina, approximately 60% of Whites supported repeal of the interracial marriage ban along with about 77% of Blacks. In Alabama, however, Whites were split about fifty-fifty on the referendum, compared to the nearly 90% support given by Blacks. Astoundingly, the racial divide in voting is comparable in size to that found during the election contest between Edwin Edwards and David Duke in Louisiana more than a decade ago.31
Duke received 55% of the White vote and virtually no Black votes (Applebome 1991).
The difference in support across the two states, particularly among Black voters, is noteworthy. We posit two explanations. First, as the GSS results (described above) suggest, voters in the South Atlantic region (which includes South Carolina) are somewhat more liberal than are those in the East South Central region (which includes Alabama). This explanation is undercut, however, by a closer examination of the GSS analysis, which reveals only negligible differences in the attitudes of Black respondents across these regions.
The second, and more likely, explanation for the difference between Black votes in each state might be in the possibility of voter confusion or error on the referenda. Since these referenda asked voters to “ban a ban,” the use of a double negative in the ballot language might have confused voters as to which vote expressed their own preference. The possibility of confusion was evident in South Carolina, where election officials included an explanation along with the text of the referendum:
AMENDMENT #4
Shall Section 33, Article III of the Constitution of this State be amended by deleting the following sentence from the Constitution: “The marriage of a white person with a Negro or mulatto, or person who shall have one-eighth or more of Negro blood, shall be unlawful and void.”
Explanation: This amendment, if approved, will remove the part of the Constitution that makes marriage between whites and blacks illegal.
In Alabama, in contrast, the text was as follows:
Proposed Statewide Amendment Number 2: Proposing an amendment to the Constitution of Alabama of 1901, to abolish the prohibition of interracial marriages.
These differences in language leave open the possibility that in South Carolina, despite (or perhaps because of) the added explanation, voters might have thought that they were being asked to approve rather than disapprove of the constitutional language banning miscegenation. In Alabama, in contrast, voters faced a shorter proposition. Confusion among South Carolina voters could also explain the fact that Blacks in that state were approximately three times as likely to vote No as Blacks in Alabama. So, perhaps voters with lower levels of education (a disproportionate number of whom are Black) were more likely to be confused by the lengthier South Carolina amendment. This idea is further borne out by evidence from a pre-election poll in Alabama. Respondents were asked the following question:
In the November general election, Alabama voters will decide whether to remove section 102 from the state constitution. The section reads as follows: “The legislature shall never pass any law to authorize or legalize any marriage between any white person and a Negro or descendant of a Negro.” As currently worded, would you say this section prohibits interracial marriage in Alabama or not?
As Table 4 shows, most Alabamans understood the ballot language, though Blacks were more likely to misunderstand it than were Whites.
Understanding of the Alabama ballot language
Regardless of whether the South Carolina results were affected by ballot confusion (with a much greater impact on Blacks), the fact remains that a strikingly high percentage of Whites in both states voted No on the proposals, which is many more than would be expected on the basis of opinion surveys. Surprisingly, at the dawn of the twenty-first century, approximately one-half of all White voters in Alabama supported a symbolic ban on interracial marriage.
This gap between survey and voting behavior also differs significantly from what would be predicted based on Krysan's recent results, in which she finds that responses to the intermarriage questions (unlike some of the other measures of traditional prejudice) showed no mode effects (Krysan 1998, p. 522). On the other hand, this finding reinforces Krysan's central argument, that context is critically important in evaluating racial attitudes.
The gap between survey and voting behavior also raises questions for Gomez and Wilson's (2006) recent claim that most of the effect of “symbolic racism” on racial-policy attitudes is eliminated by controlling for attribution bias. Gomez and Wilson hypothesize that politically unsophisticated respondents ascribe responsibility for racial inequality to individual failings rather than social causes. On its face, this is simply implausible as an explanation of attitudes toward racial marriage, and so cannot be responsible for White opposition to it.
While the available data, being essentially quasi-experimental, do not provide a definitive explanation for this gap, these results do raise the possibility that the traditional racism that defined White attitudes prior to the civil rights era is not completely dead, as some have suggested (Kinder and Sanders, 1996, pp. 92–98). It also raises the possibility (especially in view of the contrast with Krysan's findings above) that some racial attitudes may be expressed in actions that are not admitted to, even in anonymous surveys.32
In this respect, our research buttresses the findings of scholars like Keith Reeves, who have shown that White voters often mislead survey interviewers about their racial conservatism, particularly in pre-elections polls of campaigns involving with Black candidates (Reeves 1997). More generally, in the privacy of the voting booth, White voters have almost always expressed high levels of racial conservatism.
Some might also claim that opposition to these amendments, in fact reflects some form of the “new” racism, or at least something other than the “old” racism. According to such a view, those Whites who voted No did so not because they actually supported miscegenation bans, but because they saw the referendum as an example of Blacks' propensity for dredging up the past and focusing on the perceived racism of Whites. Controversial public expressions of this sort of “new” racism are not uncommon in many debates on racial matters. Thus, one would expect that if the “new” racism underpinned much of the White opposition to the measure, at least some public figures would have articulated such arguments. Yet, as previously mentioned, opposition to the amendment was virtually nonexistent, with exceptions found only among the most marginal elements. Moreover, even if White voters were motivated by something other than the “old” racism, they still expressed those motivations through supporting a legal provision that is clearly a remnant of the old racism. That many chose to do so suggests that White racial behavior is more complex and dependent on context than many have previously thought.
Of course, the applicability of this pattern of behavior beyond southern states such as South Carolina and Alabama is open to question. Support for antimiscegenation laws was, and is, higher in the South than in the rest of the nation. As a result, if Whites were reluctant to express their true racial sentiments, or if their attitudes were dependent on context in more complex ways, it would likely be more evident in this region than in others. On the other hand, we can find no reason to predict that the existence of a differential between behavior in the polls and at the voting booth is unique to the South.
CONCLUSION
The large racial division in voting on these symbolic amendments stands in contradiction to received wisdom. At the same time, it emphasizes the importance of context in evaluating attitudes. The sizeable gap between the attitudes expressed in the voting booth and those expressed in surveys suggests that evaluations of racism and racial attitudes based on survey data could be substantively incomplete. Our findings point to the desirability of incorporating information from other expressions of White attitudes in order to accurately gauge their depth and breadth.
APPENDIX
Table A.1 shows the percentage voting No, Yes, and Roll-off (abstaining from vote on that provision) as estimated using different methods: Homogenous precincts (HP), Goodman's Regression (weighted), and King's two-stage EI model. Roll-off was imputed from separate estimates of No and Yes, using a separate run to calculate standard errors. In Jefferson County, data for registration by race was available, but turnout by race was not. So, this county was estimated separately from the rest of Alabama. Since homogenous precinct analysis requires precincts that are nearly all Black or White, there are fewer available cases.
Model Estimates
