The Two Faces of Public Opinion

THE TWO FACES OF PUBLIC OPINION1 Adam J. Berinsky University of Michigan Department of Political Science e-mail: berinsky@umich.edu April 1998 1 Paper presented at the Annual Meeting of the Midwest Political Science Association, April 24-26, 1998, Chicago, IL. For many helpful discussions regarding this project and comments on earlier drafts of this paper, I would like to thank Nancy Burns, Fred Cutler, Paul Freedman, Kim Gross, Vince Hutchings, and Donald Kinder. I, of course, am responsible for any errors that remain. The data used in this paper were made available by the InterUniversity Consortium of Political and Social Research. Neither the collector of the original data nor the consortium bears any responsibility for the analyses or interpretations presented here. This material is based upon work supported under a National Science Foundation Graduate Fellowship. Any opinions, findings, conclusions, or recommendations expressed in this paper are those of the author and do not necessarily reflect the views of the National Science Foundation. In Voice and Equality, Verba, Schlozman and Brady (1995) convincingly argue that traditional forms of political participation, such as campaign volunteerism and contact with representatives, magnify the voice of the resource rich at the expense of the less privileged. There is, however, one form of political participation that would appear, at first blush, to defy the pattern of bias described by Verba, Schlozman, and Brady — public opinion polls. As Verba (1996) writes, “sample surveys provide the closest approximation to an unbiased representation of the public because participation in a survey requires no resources and because surveys eliminate the selection bias inherent in the fact that participants in politics are self-selected.” This intuition has proven to be valuable because public opinion polls have become a significant source of information for elites about what the “public” thinks about the issues of the day (see Brehm 1993, Herbst 1993; for a more cautionary note, see Ginsberg 1986). Polls are no longer the sole province of academics and journalists. Politicians and interests groups are now both producers and consumers of political polls. Public opinion surveys then are not just a form of participation in theory; the large sums of money politicians are willing to pay to underwrite polls indicates that they are being actively consumed by the political elite. But while public opinion polls may appear to be a more inclusive form of representation than traditional forms of political participation, there are a number of problems and limitations inherent in viewing polls as a source of unbiased mass input into the policy process. Before fully embracing the notion of aggregate public opinion as an unbiased form of participation, then, it is critical to conceptualize what types of signals can be sent from the mass public to elites through polls, and what messages might be distorted in that process. In some cases, aggregate public opinion suffers from the same biases as traditional forms of participation. In previous work (Berinsky 1997), for example, I found that the aggregation of social welfare policy opinions inadvertently disenfranchises from the collective opinion signal those individuals predisposed to support the welfare state. In particular, I found that the same factors which lead individuals to take a liberal position on social welfare policies — namely the lack of political resources such as political information, education, and income — also lead them to abstain from those questions altogether. This process ultimately lends a conservative bias to aggregate public opinion. My previous work, then, demonstrated that opinion polls may paint an inaccurate picture of the public’s sentiments because some people can not answer survey questions. The story, however, does not end there. Opinion polls may also be a poor reflection of collective public sentiment because other individuals do not want to answer survey questions on socially sensitive topics. Specifically, people may bow to the social pressures of the survey interview and choose to abstain from opinion polls rather than give opinions which might paint them in an unfavorable light. In the aggregate, these tendencies could have serious consequences. If significant potions of the population are loath to disclose views which could be construed as socially unacceptable, polls measuring collective opinion on sensitive topics may underestimate the true levels of support for or opposition to those policies. In this paper I trace out the aggregate effects of the social forces in the survey interview that might influence the opinions which individuals express. First, I advance the “Mediated Communication” theory of the survey response, which builds on existing models of public opinion in the political science literature by accounting for effects related to the social context of the survey setting. I then discuss how the aggregation process could compound these individual-level effects to create an opinion signal which is a poor representation of the collective public’s policy preferences. As an illustration of these effects and the resulting difficulties involved in measuring aggregate opinion on socially sensitive issues, I use National Elections Study (NES) data from 1990-1994 to show that public opinion polls overstate support for school integration. Specifically, individuals who harbor anti-integration sentiments are likely to hide their socially unacceptable opinions behind a mask of indifference. Finally, in order to confirm the validity of these 1 findings, I show that the same methods which predict that opinion polls understate true opposition to government involvement in school integration also predict the results of the 1989 New York City mayoral election — an election where the charged racial atmosphere made accurate polling difficult, if not impossible — more accurately than the marginals of the pre-election polls taken in the weeks leading to the election. All told, these results suggests that survey questions on school integration — and more generally questions on racial attitudes — may provide an inaccurate picture of true public sentiment on such sensitive issues. UNDERSTANDING THE MICROFOUNDATIONS OF PUBLIC OPINION: POLITICAL ATTITUDES AND THE SURVEY RESPONSE Stimson (1991) argues that when contemplating the place of public opinion in the political world, we must consider the collective preferences of the mass public (see Page and Shapiro 1992, Stimson, MacKuen and Erikson 1995). While this position may be accurate, it is not enough to simply focus on aggregate-level analysis; to understand the nature of aggregate opinion we must first understand the behavior of individuals who make up that aggregate. After all, individual decisions to offer or withhold opinions in the survey setting ultimately determine the shape of the aggregate signal. My analyses in this paper, therefore, proceed from a two-stage model of the way in which individuals come to first form and then express opinions on political matters. I then consider the implications of this individual-level model for our understanding of aggregate public opinion. The New Look in Public Opinion Research: Constructed Attitudes and the Survey Response Conventional theories of public opinion have treated the survey response as the product of individuals’ attempts to reveal their fixed preference on a given policy issue. Controversies in the literature, therefore, have traditionally centered on how well the survey instrument enables people to translate their preferences into survey responses and whether the over-time attitude instability identified first by Converse (1964) was due to vague minds (Converse 1964) or vague questions (Achen 1975). Recently, however, a more fluid view of the survey response has emerged, based in part on theories of preference construction developed in cognitive psychology. This view, advanced most forcibly by Zaller and Feldman (Zaller and Feldman 1992; Feldman 1989; Zaller 1992), argues that “individuals do not typically possess ‘true attitudes’ on issues, as conventional theorizing assumes, but a series of partially independent and often inconsistent ones.” (Zaller 1992, p. 93). According to this “new look” in public opinion research, a survey response is not necessarily a “revealed preference.” Instead, answers to survey questions should be viewed as a realization of a stochastic draw from an individual’s underlying response distribution, which itself is an aggregation across one’s potentially diverse feelings — or “considerations” — concerning political issues. Thus, survey responses do not reveal a single “true attitude,” but instead reveal a sample of the types of concerns and considerations people bring to bear when considering issues in the realm of politics. This conception of the survey response has proven to be a powerful way of thinking about public opinion. For example, Zaller and Feldman (1992) demonstrate that their model can explain not only the over-time instability in some individuals’ response to identical survey questions but also the “framing effects” of question-wording and question order.1 But while Zaller and Feldman’s model is powerful and represents a significant conceptual 1 I use Zaller and Feldman’s model as a reference point here because it is the most fully specified political science model of the survey response. Zaller and Feldman’s work, however, bears a strong resemblance to many other conceptions of the survey response (Iyengar and Kinder 1987, Kinder and Sanders 1990, Bartels 1988, Chong 2 advance over the (sometimes implicit) assumption that individuals have singular “true attitudes” that they reveal to survey interviewers in a direct and unbiased manner, the model is incomplete. Specifically, it does not account for response effects arising from the social nature of the interaction in the survey interview. Putting the “Social” into Survey Research Zaller and Feldman’s focus on the cognitive processes of opinion formation echoes early psychological models of the survey response which presented straightforward extensions of general information processing models developed in cognitive psychology (for a brief review, see Sudman, Bradburn, and Schwarz 1996). Like those models, Zaller and Feldman’s work concentrates on the individual’s thought processes at the expense of a consideration of the social context in which the survey interview takes place.2 Ignoring the social context, however, omits a key factor from consideration. The interview, after all, is a form of social interaction between two individuals; the interviewer and the respondent. Thus, as Sudman, Bradburn, and Schwarz (1996) note, the survey interview is: best considered as an ongoing conversation in which respondents conduct their own share of thinking and question answering in a specific social and conversational context. Hence, conceptualizations of the question answering process in survey interviews need to consider conversational as well as cognitive processes and pay close attention to the complex interplay of social communication and individual thought processes. (p.55.) Survey researchers have long recognized that the environment of the survey setting could have potentially significant effects on the nature of the opinions individuals express (Hyman 1954).3 To use Hyman’s words, it has long been acknowledged that “attitudes are not independent of the circumstances within which they are liberated.” For example, Post-War studies of public opinion found that interviewers purporting to represent the (fictitious) German Opinion Institute were less likely to obtain valid opinion answers from respondents than were unaffiliated interviewers (see Hyman 1954). With these concerns in mind, psychologists have proposed a number of models of the survey response which integrate the cognitive and social dimensions of both opinion formation and expression (Tourangeau and Rasinski 1988, Wilson and Hodges 1992, Tourangeau 1992, Sudman, Bradburn, and Schwarz 1996). These models use somewhat different terminology to describe the various tasks involved in the question-answering process, but they all offer a more complete view of the survey response process than that offered by Zaller and 1993; 1996) and is, therefore, “standing for” those models in this discussion. 2 To be fair, Zaller and Feldman acknowledge that cues in the social environment — such as the race of the interviewer — could affect the survey response, but only insofar as people might pick up “cues” from that environment. Specifically, people might be cued by the immediate social context — either consciously or subconsciously — to give greater weight to particular considerations when they are unsure of their opinion. Zaller and Feldman specifically say that social desirability — “where people consciously mis-report attitudes in order to avoid embarrassment” — is at odds with their interpretation of the survey response. 3 The preoccupation with the social realm here, it should be noted, is somewhat different from the social concerns advanced by scholars critical of the “atomistic” tradition of opinion collection and analysis. Beginning with Blumer (1948), a number of scholars have argued that public opinion is formed through the interaction of individuals and groups of individuals. These authors argue that care must be paid to the social context of individual. Otherwise, as Blumer notes, opinion research will fail to capture opinions as they are organized and operate in a functioning society. This line of work continues to the present day, most recently and persuasively by Huckfelt and Sprague (1995). But while this line of work is extremely promising, I argue that a concern for the effect of the broader social context of society must not come at the expense of a rigorous examination of the social context of the interview. Specifically, we must pay attention not only to the effects of social networks on the transmission of information, but also the effects of social relationships on the nature of attitude expression. 3 Feldman. Specifically, these models recognize that the translation of an individual’s private summary judgment into a survey response is accomplished by means of a social interaction between the survey respondent and the interviewer. For the present purposes, the important point to draw from this literature is that the social context of the interview setting can influence the survey response and — under some circumstances — the respondent may not faithfully reveal their private judgment to the survey researcher. The Mediated Communication Model of the Survey Response It is clear, then, that the Zaller and Feldman model of the survey response, while powerful, is somewhat incomplete. To provide a more complete conception of the survey response, then, I expand on the Zaller and Feldman model by integrating elements of the psychological work discussed above into a “Mediated Communication” (MC) model of the survey response (see Berinsky 1998). The MC model posits that respondents answer survey questions in two stages. In the attitude formation stage, the respondent constructs a summary evaluative judgment of whatever it is they are being asked to evaluate — In technical terms, the “attitude object.” In the second opinion expression stage, the respondent relays that judgment to the interviewer in the context of the survey setting. Like the Zaller and Feldman model, the MC model is a representation of the process by which respondents construct and communicate their preferences through surveys. But my model offers a more comprehensive conception of the survey response than do those authors by incorporating response effects which arise from the fact that this communication is mediated through the social interaction of a survey interview. Under many circumstances, it is safe to presume that respondents will communicate their attitudes in a straightforward and unaltered manner. However, the social nature of the interaction between the respondent and the interviewer might thwart such a candid conveyance of a respondent's wants, needs, and desires. In sum, under some circumstances, the opinion constructed by the respondent is not necessarily the same as the opinion expressed by that respondent. This disjunct between private opinion and public utterances could result from several factors. To begin with, respondents could misconstrue their preferences because of “social desirability effects” which arise from a respondent’s desire to hide attitudes that society as a whole might deem “unacceptable.” Singer, Von Thurn and Miller (1995) find, for example, in a broad review of the literature, that the balance of experimental evidence indicates that significantly different response are often obtained on sensitive questions as the privacy of the response process is increased. These results suggest that respondents’ public opinion may not faithfully reflect their private attitudes. This phenomenon is especially prevalent in questions concerning racial matters. For example, Dovidio and Fazio (1992) find that expressions of racist principles increase as respondents believe the truthfulness of their statements are somehow being monitored. Similarly, Hurley (1997) finds that experimental measures which unobtrusively measure racial prejudice reveal racism to be far more prevalent than do traditional measures of racism — which are presumably tainted by the residues of the respondents’ efforts to give socially desirable answers. Thus, social costs could lead to individual reticence, or even silence, on potentially embarrassing matters.4 4 This concern with social desirability, broadly defined, has even been extended to the study of preferences in an economic framework. Kuran (1995) argues that seeking to maximize one’s reputational utility — the payoff a person receives from public responses to their private preferences — could, and sometime does, lead individuals to take public positions at odds with their private beliefs. 4 Most authors in the survey research literature argue that these social desirability effects are mediated through the specific conditions of the survey interview — in particular in the relationship between a specific interviewer and a particular respondent. These scholars argue that respondents are sensitive to group-specific social cues in an interview, such as the race and perceived social class of the interviewer (Hyman 1956). The “race-of-interviewer effects” identified by numerous scholars (see Anderson, Silver, and Abramson 1988, Davis 1997, Finkel, Guterbock, and Borg 1991, Kinder and Sanders 1996, Schuman and Converse 1971) fall into this class of effects. As Finkel, Guterbock, and Borg note, “the most common explanation for [interviewer] effects is that individuals seek to avoid offending interviewers of the opposite race because of ‘interpersonal deference,’ ‘courtesy to a polite stranger,’ or more simply a general desire to be agreeable” (p. 315). While the literature concerning race-of-interviewer effects is important and interesting in its own right, it is important to recognize that in addition to these interviewer-specific effects, social desirability effects could arise from the very nature of the social interaction of the survey setting. In other words, broad concerns that transcend the specific social transaction between the interviewer and the respondent could also alter the nature of the survey response. For example, Schuman et al. (1997) find that “at least some white respondents feel pressure in an interview situation to appear more racially liberal than they would indicate under conditions of greater privacy,” even when the interviewer is also white (p. 95). Social desirability effects may, in part, depend on who is asking the question, but pressures to hide attitudes that might paint the respondent in an unfavorable light may not be simply a function of race-of-interviewer effects. Thus, more general social desirability effects may arise from the broader political and social milieu in which the survey conversation takes place. However, regardless of the source of social desirability effects, it is plausible that under circumstances where respondents fear that they might be “censored” or socially shunned for their attitudes — either by society or the interviewer — they might shade their attitudes when reporting them to the interviewer.5 Considering the “Don’t Know” Response: Opinionation and Action The MC model of the survey response is generally applicable to understanding how people answer survey questions. But in this paper I consider a particular class of survey-related decisions. Specifically, I examine the respondent’s decision whether or not to answer specific survey questions. The reason for my interest in this question is straightforward; when considering the composition of the “public voice,” as measured in public opinion polls, it is critical to contemplate just whose voice might be lost or muted in the process of aggregation. The most obvious way that an individual’s voice could be lost in the aggregation process is if they choose to abstain from a particular question.6 The MC model suggests that individuals come to the “don’t know” answer by one of two very different routes; either when they first form their opinion about a particular political controversy or when A word of caution is in order here. When discussing social effects, many researchers discuss only the factors that might prevent an individual from speaking what is on their mind. While it is important to consider the social costs of opinion expression, it is also critical to consider the benefits a person receives from revealing their underlying sentiments on a particular political matter. Not all individuals are equally attentive to social concerns (Snyder 1983). In fact, some individuals might incur personal costs by not expressing their private preferences. Kuran (1995) posits that individuals might seek to maximize not only their reputational utility (as described above), but also their expressive utility. Expressive utility, he writes, is the satisfaction an individual receives by supporting publicly their private beliefs. We need, therefore, to attend to inter-individual heterogeneity in the social costs to different individuals. A social cost that might be too burdensome for one might be insignificant for another. 6 This analysis, of course, assumes that respondents have agreed to participate in the survey itself (see Brehm 1993 for a detailed discussion of the causes and consequences of survey non-response). 5 5 then express that opinion to the survey researcher. Most models of the decision to abstain from survey question consider only the first option (Schuman and Presser 1981; Krosnick and Milburn 1990). Such models assume that respondents abstain from survey questions because they “don’t know” how they feel about an issue, policy, or political candidate. This is certainly true in many cases, but it is also possible that individuals could arrive at a “don’t know” answer because they do not wish to reveal their political preferences to the survey interviewer. If a person constructs a private judgment that they do not feel comfortable expressing — for example, if they feel that their preferred answer might cast them in a poor light — they might choose to abstain from the question, rather than revealing an embarrassing opinion or shading that opinion to fit what they believe would be socially acceptable behavior. Given that individuals are willing to shade their opinions when reporting them to survey interviewers, as described above, it should not be surprising that they might choose to withhold socially unacceptable views altogether. The “spiral of silence” theory (Noelle-Neuman 1984) in fact predicts that individuals who hold viewpoints that they perceive as being in the minority will remain silent for fear that expressing their views will lead to social isolation.7 While the evidence on this theory is mixed at best, a weaker view seems acceptable. Because survey responses are low benefit acts, then if the respondent is uncomfortable with their view, then they may say that they “don't know” how they feel about a particular issue rather than incur even minimal social costs. AGGREGATE EFFECTS: SCHOOL INTEGRATION Certainly, response effects related to social concerns will not contaminate the measures of opinion on all issues. In order for individual-level social effects to bias aggregate measures of public opinion, the social context of the survey setting must systematically affect the expressed opinions of individuals — or at least a group of individuals — in the sample. To the extent that individual characteristics systematically affects the willingness or ability of some  if not all  respondents to answer survey questions, the aggregate signals measured in public opinion surveys may paint a distorted picture of underlying collective public sentiment. It is important, then, to identify specific issue areas where we might find such systematic social effects. One important dimension that scholars have considered when assessing the validity of aggregate signals sent by the public to elites through opinion polls is the “hard” versus ”easy” issue dimension identified by Carmines and Stimson (1980). Those authors argue that some issues are “hard” in the sense that they require careful consideration of technically difficult choices relating to the means by which government should respond to (often) novel problems on the policy agenda. “Easy” issues on the other hand, involve symbolic concerns relating to the ends of public policies long in the public’s eye. But while the (largely) formative complexity dimension captured by Carmines and Stimson is undoubtedly an important factor to consider when gauging the validity of public opinion polls as a representation of how the public feels about the issues of the day, that typology is ultimately incomplete. Issues may, after all, be “hard” and “easy” not only in the cognitive sense identified by Carmines and Stimson, but also from a social standpoint. Thus, analysts and consumers of opinion polls must also pay attention to the ease of opinion elicitation in particular issue areas. As noted above, work in psychology suggests that the measurement of attitudes concerning sensitive topics may be hindered by respondent’s attempts to give 7 It should be noted that, unlike Noelle-Neuman’s theory, the MC model does not presume that the majorityminority status of a respondent’s opinion necessarily affects that respondent’s willingness to express their opinion. In fact, respondents who perceive they are in a majority may choose to keep their views private if they believe that the social cost of expressing those majority views is prohibitive. 6 “socially desirable” statements (Renzetti and Lee 1993; for applications in political science, see Jackman 1997 and Hurley 1997). This tendency might provide just the type of systematic individual-level effects that could lead to faulty aggregate public opinion measures. One area of politics that seems to contain a good number of cognitively easy, but socially difficult, issues are questions relating to policies of racial equality. Over the past 40 years, a sea-change in racial attitudes has occurred. Segregationist and exclusion principles that were accepted by a large majority of Americans in the 1940s and 1950s find bare traces of support today (see Schuman, Steeh, Bobo, and Krysan 1997). But while overall support for the position of racial equality has increased greatly since the middle of the century, many Americans remain profoundly ambivalent about polices which seek to use government efforts to improve the position of blacks. Thus, while many Americans may support general principles of equality, large segments of the mass public are more ambivalent about programs designed to implement those policies (Jackman 1978, Schuman, Steeh, Bobo, and Krysan 1997). Under such circumstances, respondents may be inclined to hedge and moderate their racially conservative views on policies to avoid appearing to be subscribe to racist principles. Previous work suggests that the level of ambivalence and hostility to racially liberal policies may be, in part, impacted by the social environment of the interview, as described above. For example, Schuman, Steeh, Bobo, and Krysan (1997) find that attitudes in a racial survey administered by mail appear more racially conservative than attitudes obtained in a more tradition face-to-face survey setting. As the authors note, this result indicates that at least some (white) respondents feel pressure in an interview setting to give more racially liberal responses than they, in fact, ascribe to (see also Dovidio and Fazio 1992, Hurley 1997, Kinder and Sanders 1996). This trend seems to extend to reports of behavior. Preelection polls of registered voters in electoral races where one candidate is white and the other black usually overestimate support for the black candidate. The social pressures of the survey interview may even, in some instances, drive ambivalent and racially conservative respondents to take a “don’t know” response. For example, pre-election polls in bi-racial elections typically amass an unusually large number of “don’t know” responses. As some scholars have noted (see for example Reeves 1997), these “don’t knows” may, in fact, provide a socially acceptable cover for support for the white candidate. INSERT TABLE 1 ABOUT HERE Questions which tap into racial policy concerns, then, would seem to provide a fertile arena in which to test my theories concerning the effect of the social pressures of the survey interview setting on expressed public preferences. One question which is particularly well suited to my purposes is the NES school integration question, which asks if respondents support government intervention to ensure that black and white children go to the same school. This question is advantageous because while the American public overwhelmingly supports the principle of school integration, they are ambivalent about the use of government authority to ensure that integration occurs. In the 1990-1994 period, a little less than half of the respondents supported government intervention (see Table 1). More important for the present purposes, however, is the fact that almost one-third of the respondents chose to abstain from the question in each of those years. This unusually high number of “don’t know” responses may, in part, be a function of the fact that NES uses a “full filter” (Schuman and Presser 1981) for the integration question. Specifically, respondents are first asked if they have an opinion on the integration question, and then are asked what that opinion is.8 This structure gives respondents an opportunity to opt out of the question-answering process at 8 The question reads, “Some people say that government in Washington should see to it that white and black children go to the same schools. Others claim that this is not the government’s business. Have you been concerned (1990: interested) enough in this question to favor one side or the other? [If yes] Do you think the 7 an early stage and — quite possibly, if my hypothesis is correct — avoid a socially uncomfortable response by saying that they “don’t know” how they feel about the integration question.9 MODEL SPECIFICATION: SCHOOL INTEGRATION AND SELECTION BIAS If my intuitions about the school integration data are correct, then a number of respondents who oppose government involvement in school integration might choose to abstain from the integration question rather than reveal an opinion that could potentially be construed as socially unacceptable. If this is the case, the sample of individuals who offer an opinion on the integration question will differ in systematic ways from the sample of non-respondents. Put another way, if some respondents who oppose integration hold their tongues rather than voice their opposition, those individual who do answer the integration question will be “unusual” in the sense that they are more likely to hold favorable attitudes towards school integration than the rest of the population. Thus, if my hypothesis is correct, the school integration question-answering process will be plagued by a selection bias problem. 10 Achen (1986) notes that the effects of selection bias can be avoided in regression analysis if and only if the unobserved factors influencing selection are uncorrelated with the unobserved factors influencing outcomes.11 Such a state of affairs will arise here only if the process by which people decide if they have government in Washington should see to it that white and black children go to the same schools, or stay out of this area, as it is not their business?” 9 Before I continue, it is important to note that my argument is not that respondents necessarily choose to hide their sentiments concerning school integration behind the “don’t know” response because they do not want the world to know they are racists. Instead, some people might fear that their opposition to school integration might be construed as racist, even though their opposition might result from legitimate policy concerns. Thus, given a situation when an attitude could be construed as revealing something unpleasant — and quite possibly untrue — about one’s character, it is perfectly reasonable to expect that opponents of school integration might “pass” rather than placing oneself in a socially difficult situation. This trend could hold for blacks as well as whites. Blacks too have legitimate reasons for opposing government efforts to integrate schools. But explaining that position might be more difficult than is worth the trouble. In sum, then, hesitancy to reveal one’s true opinion is not necessarily an indicator of racism. Rather, it could simply result from a desire to avoid a situation where one’s intentions could be misconstrued as a sign of a socially unacceptable attitude. 10 While selection bias is important in this paper as a diagnostic tool, selection bias is a potentially serious problem in its own right because it can have highly deleterious effects on our inferences. Ignoring the sample selection mechanism in effect omits a variable — the effect of selection — from the outcome equation. This form of omitted variable bias is illustrated most transparently in the context of the Heckman selection model (Heckman 1979). In that instance, the outcome equation estimated by itself omits the expected value of the error term of the outcome equation under censoring — otherwise known as the hazard rate (Achen 1986). This analogy of the ignored selection mechanism to the omitted variable bias —identified first by Heckman — is transferable to other selection models, such as the Tobit (Breen 1996) or the bivariate probit selection model (Dubin and Rivers 1990). In all these cases, ignoring the sample selection mechanism in the presence of selection bias will lead to two problems. First, estimation will produce biased estimates of β because, in essence, the variable measuring λ has been omitted. More precisely, we say that the estimates of β are inconsistent (Maddala 1983). In addition, the estimates of β will be inefficient, because the error term of the outcome equation is heteroskedastic. Thus, T-tests and other classical hypothesis tests will lead to incorrect inferences concerning the statistical significance of variables in the outcome equation (Greene 1997). 11 Specifically, as Dubin and Rivers (1990) note, in the event of selection bias, errors occurring in the outcome equation sample do not have zero mean because the sampling procedure has picked out those observations that are, in terms of the theory, “unusual.” Achen (1986) makes the same point in the case where the outcome equation dependent variable is continuous (and is, therefore, estimated using OLS). Specifically, he notes that if the error 8 an opinion and the processes by which they decide their opinion are independent events.12 In other words, in the case of the school integration data, selection bias can be avoided only if the decision to offer an opinion on the school integration question is independent of one’s opinion on that question. Models which account for selection bias, then, will enable me to directly test my hypothesis concerning the effects of the survey interview setting on expressed opinion. In particular, selection models will measure the extent and nature of the link between the decision to offer an opinion on school integration and the decision to take a position on that issue. If the estimates reveal that selection bias does not exist in the school integration question-answering process, then we can say with confidence that my theory concerning the link between social desirability concerns and the “don’t know” response is incorrect.13 Conversely, to the extent that my statistical analysis reveals the presence of selection bias, we may reasonably infer that the social context of the survey interview leads some individuals to suppress their preferences and offer a “don’t know” response instead. We can then use the results of these analyses to determine the nature of the preferences concealed by the don’t know response. Dubin Rivers Model Several models exists which account for the statistical effects of the process of selection into the sample of respondents. One correction that is especially well-suited for my purposes is bivariate probit selection model, described by Dubin and Rivers (1990). The bivariate probit selection accounts for selection bias in the case of an outcome equation with a binary dependent variable. The Dubin-Rivers model is intended to analyze data, like the NES school integration question, which have been censored — that is data for which we do not have information about the dependent variable of interest for some respondents, but we do have some information about the attributes of the non-respondents (Breen 1996, Greene 1997). This model combines information about the untruncated successes and failures (the outcome equation) with information about the untruncated observations (those excluded by the selection equation) in a MLE switching model. The Dubin-Rivers model can be represented as a bivariate probit with one of the quadrants collapsed over those individuals who are selected out of the outcome equation. The log-likelihood of this model is, following Greene (1995):  Ln L (β 1 , β 2 , ρ ) = ∑ y 2 =1, y1 =1 ln Φ 2  β 1′ xi1 , β 2 ′ xi 2 , ρ     + ∑ y2 =1,y1 = 0 ln Φ 2  − β 1 ′ xi1 , β 2 ′ xi 2 , − ρ     + ∑ y2 = 0 ln Φ − β 2 ′ xi 2    (1) terms in the selection and outcome equations — u1i and u2i respectively — are correlated in the censored sample, the disturbance term of the outcome equation u 2i has neither mean zero nor zero correlation with the outcome independent variables, even though it has both properties in the full sample. Thus, when the sample selection process is related to the error of the outcome equations, separate estimation of the selection and outcome equation will lead to faulty inferences concerning the effect of the variables of interest in the outcome equation. 12 Technically, selection bias can also be avoided if every variable influencing selection is controlled in the outcome equation. But, as Achen argues, it is practically impossible to fully “control” for the selection mechanism in the outcome equation and achieve this condition. 13 By no means is it ordained that selection bias exists in the data. For example, estimates of models of responses to social welfare questions indicate that those opinion are not contaminated by selection bias (Berinsky 1997). 9 Where: Yi2 ~ fbern (y1i | π1i), π1i defined by the underlying probability term Yi 1 = β xi1 + u1i is the outcome process, Yi2 ~ fbern (y2i | π2i), π2i defined by the underlying probability term ∗ Yi 2 = β xi 2 + ui 2 , is the selection process y1i =0 and y2i =1 is an untruncated failure, y1i =1 and y2i =1 is an untruncated success, y2i =0 is a truncated observation. ∗   Φ 2  β 1′ x1 , β 2 ′ x 2 , ρ is the cumulative bivariate normal function defined by   β ′ x , β ′ x and ρ; 1 1 2 2 and u1i and u2i are bivariate normally distributed iid, with σ u1,u 2 = ρ .14 Estimation of this model yields three sets of parameters: β1, the effect parameters for the outcome equation; β2, the effect parameters for the selection equation; and ρ (rho), the correlation of the errors between the two equations. For the present purposes, it is the ρ parameter that is of particular interest because it measures the link between the selection and outcome processes — specifically, the link between the unmeasured factors in the selection equation and in the outcome equation. Here, I argue that these unmeasured effects arise from the social setting of the survey interview which lead some respondents, who are hostile towards government efforts to integrate schools to hide their sentiments behind a “don’t know” response. Therefore the presence of these social desirability effects can be gauged by the statistical and substantive significance of ρ. If ρ=0, then the unmeasured factors which influence whether a respondent will offer a response to the school integration question are independent of the unmeasured factors which lead that respondent to support or oppose government efforts to integrate schools. Under such circumstances, my theory concerning the presence and effects of social concerns in the survey interview must surely be incorrect. If, on the other hand ρ is positive, then the unmeasured factors which lead someone to answer the school integration question are correlated with the unmeasured factors which lead then to give a more racially liberal response on the integration question. In other words, if ρ is positive, then individuals who do answer the integration question are — by simple dint of answering the question — more likely to support integration than individuals who abstain from the question. We may then conclude that some individuals who harbor anti-integration sentiments are likely to hide their socially unacceptable opinions behind a mask of indifference. Thus we can say that the social setting of the survey interview leads to selection effects and — more importantly — can trace out the aggregate implications of that bias for aggregate public opinion on government policies regarding school integration. MODEL SPECIFICATION In practice, this model is identified as long as the same variables are not included in both the selection and outcome equations. That is, while it is possible to identify the model through the nonlinearity of the selection equation, identification should proceed from exclusion restrictions. 14 10 One possible strategy for gauging the presence and effects of selection bias in the school integration data would be to estimate a single model of the question-answering process. Many previous efforts in the political science literature to gauge selection bias effects have followed such a strategy (see, for example, Dubin and Rivers 1990). While such an approach is certainly valid, estimating only a single model of the question-answering process allows us only a partial view of any selection bias. Estimating multiple models of the question-answering process allows us to be sure that any correlation between the errors in the two equations we estimate is intrinsic to the question-answering process and not simply an artifact of omitted explanatory variables common to both equations.15 In the analyses that will follow, therefore, I will estimate several equations using data from the Nation Election Study (NES). These equations model, first, the process by which people decide whether they have an opinion which they are willing to reveal in the survey interview and then, second, the process by which they decide what that opinion is. Specifically, I consider two models of the process by which people decide to answer the school integration question — (1) the “engagement” model and (2) the “combined” model — and four models of the process by which the question respondents form their opinion — (1) the “values” model, (2) the “demographic” model, (3) the “self-interest model,” and (4) the “full” model.16 These equation specifications are described below. In order to consider all possible permutation of these models, I cross the two selection equations with the four outcome equations to yield eight separate models. Admittedly, this strategy may verge on overkill. But to the extent that different models of the question-answering process yield similar conclusions regarding the presence and effects of selection bias, we will gain greater confidence that these effects are endemic to the question-answering process and not simply an artifact of model mispecification. Opinionation (Selection) Equations 1. ENGAGEMENT MODEL. Previous work (Schuman and Presser 1981; Krosnick and Milburn 1990, Berinsky 1997) has found that an individual’s propensity to offer an answer on a survey is related to their general level of engagement with the political system. I, therefore, included in this first selection equation a number of variables which proxy for general political engagement, namely: the respondent’s level of political information (see Zaller 1992), a dummy variable measuring whether the respondent will place themselves on a liberal-conservative scale, and a measure of how often the respondent discusses politics. In addition, it is possible that respondents who will be affected by school integration policies will be more likely to be engaged with that topic and, by association, will be more likely to have an opinion on the integration issue. Thus, I also included variables measuring the respondent’s race and the number of children they had in school.17 Finally, I included three measures indicating how difficult it was to contact As Breen (1996) notes, “the correlation [between the selection and outcome equations] should be thought of as intrinsic to the model. In other words, we assume ρ ≠ 0 in the theoretic model that we posit for the population and not simply for the sample in which we may have omitted the measurement of a variable common” to the selection and outcome equations. The errors, therefore, covary despite proper model specification. The cause of the correlation terms should, then, be unmeasurable. “In essence,” Breen concludes, “both equations are affected (in part) by the same random perturbations (or random perturbations that tend to covary)” (p. 35.). 16 I initially attempted to estimate a third selection equation, which consisted of the same variables as the outcome equation “demographic” model. I abandoned this effort when it became clear that the effect of the administrative selection variables — whether the respondent was sent a refusal conversion letter, whether the respondent received a persuasion letter, and the number of calls required to secure the interview — were not strong enough to satisfy the bivariate probit model exclusion restrictions. 17 To capture possible ceiling effects on the power of this variable, I included the natural log of the number of school-age children in a family, rather than the untransformed form of that variable. 15 11 the respondents (Brehm 1993) — whether the respondent was sent a refusal conversion letter, whether the respondent received a persuasion letter, and the number of calls required to secure the interview — on the assumption that those who are difficult to reach would also be reluctant to answer specific survey questions. 2. FULL MODEL. While the engagement model is a more than reasonable predictor of whether a respondent will offer an opinion on school integration, It might be that the engagement model does not capture all the effects related to the decision to express a preference on the integration question. Thus, I developed a second selection equation that expanded the engagement model by including a number of demographic variables. Specifically, I included measures of a respondent’s: age, race, sex, education, income, area of residence, religion, occupation, and home ownership status. The addition of these variables may, admittedly, not provide a complete model of the process by which people decide to answer the integration question. But, following Bartels (1996), with the addition of the demographic variables, this equation “might be thought of as a reduced form model corresponding to some more elaborate, more realistic, but unspecified structural model in which proximate political attitudes mediate between demographic and social characteristics on the one hand” and the decision to offer an opinion on the other (p. 208). Opinion (Outcome) Equations 1. VALUES MODEL. In the first outcome equation, I estimated the mean (Xiβ) as a function of measures of general political affiliations and political principles. My primary measure of a respondent’s political affiliation was partisan identification.18 To gauge the effect of political principles, I included dummy variables indicating self-identification — or non-identification — with an ideological group.19 However, such classifications by no means exhaust the possible predictive power of political principles. While affiliation with these broad ideological categories are almost certainly important determinants of opinion direction, as over thirty years of public opinion research have clearly demonstrated, “the political thinking of much of the public cannot be adequately described as ideological in the sense of deductive reasoning from an overarching set of integrated principles about politics and the social world.” (Feldman 1988, p. 417; see also Kinder 1983). In addition to — or instead of — ideological self-identification, people use more diffuse principles — what Feldman (1988) terms “core beliefs and values” — to understand the political world (see also Feldman and Zaller 1992; Kinder and Sanders 1996; Kinder 1998). Thus, to capture these “sub-ideological” grounding principles, I included the respondent’s level of trust in government and measures of support for equality, racial resentment, and moral conservatism. Finally, I included measures of respondents’ levels of religious engagement to disentangle “strong traditional religious conviction” from the “moral conservatism” concept in which I am interested. 2. DEMOGRAPHIC MODEL. The second outcome equation posits that a respondent’s position on the school integration question can be modeled as a function of their demographic characteristics. Specifically, the demographic model includes measures of a respondent’s: age, race, sex, education, income, area of residence, religion, occupation, and home ownership status. As with the “full” selection model, this equation may be viewed as a “reduced form” model of a more complex model of the choice process. 3. SELF-INTEREST MODEL. In the third outcome equation, I estimated the mean (Xiβ) as a function of measures of respondent’s self interest in the school integration issue. Specifically, as in the first selection For the purposes of these analyses, I have collapsed party identification to a five point scale from the traditional seven point scale (see Keith, et al 1992). 19 The excluded group was those individuals who identified themselves as politically “moderate.” 18 12 equation, I included variables measuring the respondent’s race and the number of children they have in school. Finally, I included measures of respondent income under the assumption that lower income respondents would have more stake in the question of public school integration than higher income respondents because higher income respondents could conceivably send their children to private school and avoid the school integration question altogether. 4. COMBINED MODEL. My last model included all the variables contained in the first three models. This equation therefore models school integration opinion as a function of general political affiliations, material interests, and political principles (See Kinder 1998). More importantly for the present purposes, it presents a comprehensive model of the process by which people determine their opinion on school integration and, therefore, reduces the chances that any selection bias estimated in the models would be an artifact of specification error. INSERT TABLE 2 ABOUT HERE MODEL ESTIMATION AND INTERPRETATION Table 2 presents my coefficient estimates for one model of the complete school integration question-answering process. In this model, I use the “engagement” equation to model selection into the sample of question-answerers and the “demographic” equation to model choice on the school integration question. I present coefficients for both (1) the case where the outcome equation is estimated separate from the selection equation — the “independent probit” model column — and (2) the case where the selection and outcome equations are estimated jointly in the bivariate probit selection model — the “bivariate probit” column — thereby accounting for possible selection bias in the question-answering process. As the table demonstrates, both the selection and outcome equations contain a number of statistically and substantively significant predictors of opinionation and attitude choice. The more interesting matter for the present purposes, however, is whether those individuals who volunteer an opinion on the school integration question differ systematically from the non-placers. It is this question, after all, which will determine whether my hypothesis concerning the effects of the social milieu of the survey interview are correct. The short answer to this key question is an unequivocal “yes.” As Table 2 shows, the coefficient on ρ is highly significant in both a statistical and a substantive sense. Thus, it appears that the process by which individuals decide to offer an opinion is not independent of the process by which they decide what that opinion is. Because ρ is positive, this result suggests that, as I hypothesize, the unmeasured factors which lead someone to reveal their answer to the survey interviewer also lead them to take a more supportive stance on the integration issue. Conversely, individuals who harbor anti-integration sentiments are likely to hide their opinions behind a mask of indifference. Identifying the presence of selection bias, though important, is only half the story. Ultimately, what is important is ascertaining the effects of that bias. Correcting the outcome equation for the selection bias present in the question-answering process alters the model estimates in several ways. A comparison of the independent probit results and the bivariate probit results indicates that using the Dubin-Rivers correction alters the substantive performance of many outcome equation variables. While none of the substantively significant coefficient estimates change sign, the substantive power of many of the coefficients are altered by the selection bias correction. For example, the coefficient on the variable indicating whether the respondent is Hispanic is attenuated by 24 percent once the correction for selection bias is undertaken. This result indicates that Hispanics are not as distinct from whites in their opinions on 13 school integration as the separate probit estimate would suggest. The movement of the coefficient estimates across the two model specifications is certainly significant, but what is more impressive is the movement of the constant term. While the intercept in the school integration choice outcome equation is approximately zero in the separate probit model, it is highly negative in the bivariate probit selection model. This result indicates that once selection effects are accounted for, everyone is more likely to oppose school integration than the separate probit model suggests. In effect, correcting for selection bias attenuates the probability of supporting government efforts to integrate schools in the aggregate. INSERT TABLES 3-5 ABOUT HERE The results of the other seven models clearly demonstrate that the selection bias identified in Table 2 is not an artifact of omitted variables. Rather the bias is endemic to the question-answering process. The coefficient estimates for two of the other seven model runs are presented in Table 3 and Table 4. As the tables show, the selection bias effects demonstrated in Table 2 hold even when the selection equation is expanded to included demographic terms (Table 3) and when — in addition — the outcome equation is expanded to include a variety of demographic and attitudinal variables (Table 4). In both cases, the coefficient on ρ is highly significant in a substantive sense and — in the case of the model presented in Table 3 — in a statistical sense as well. More important for the present purposes, the general trend of lowered support for school integration once selection bias is accounted for in statistical estimation endures across other specifications. As Table 5 demonstrates, the basic pattern found in the three models presented above — a significant ρ and a reduced constant term once the correction for selection bias is introduced — hold across all eight models estimated.20 INSERT TABLES 6-7 ABOUT HERE Having demonstrated the presence and effects of selection bias in the school integration questionanswering process using the 1994 NES data, I sought to replicate my results using data from 1992 and 1990. Tables 6 and 7 present the coefficient estimates for the “comprehensive” model — the full selection equation and the combined outcome equation — using data from those two years.21 This quasi-experiment is the most stringent test of my hypothesis concerning the effects of the social environment of the survey interview because this model takes into account the greatest variety of demographic and attitudinal variables of any of my eight models. Thus, we can be confident that any observed selection effects are “real” rather than a statistical artifact of omitted variables. The complete model results are available from the author upon request. The 1990 sample is much smaller than the 1994 and 1992 samples because only one-half of the 1990 sample was asked the school integration question. Because respondents were randomly assigned to the form with the school integration questions, it is valid to draw inferences from that sample to the full population. 21 20 14 As the tables demonstrate, the selection effects found in the 1994 data are certainly present — and in fact may be stronger — in the 1992 and 1990 data.22 Not only is ρ highly significant in both a statistical and a substantive sense, but the constant is reduced greatly once selection bias is accounted for. Both of these results are consistent with the pattern found in the 1994 data. In addition, a trend that was suggested in the 1994 analysis comes through strongly in the data from the earlier surveys. In both 1990 and 1992, correcting for selection bias significantly increases the size of the coefficient on the dummy variable indicating whether the respondent is black. In particular, the coefficient on the “black” variable increases by 22 percent in 1992 and by 57 percent in 1990 relative to the separate probit equation. This result indicates that the gap between whites and blacks in support for school integration may be even larger than an analysis of the sample of the respondent who offer an opinion on integration suggests.23 This difference aside, these model estimates clearly demonstrate that the result which replicates across the eight different models in the 1994 data also replicates across time. The process by which people come to express their attitudes on the school integration question, then, is certainly contaminated by a selection bias which conceals significant anti-integrationist sentiments behind the veneer of a “don’t know” response. Race-of-Interviewer Effects Though the comprehensive model specification — the full section model and the combined outcome model — contains a complete set of demographic and attitudinal variables, it is possible that the specification is incomplete in that it does not account for the specific conditions of the survey setting and the respondent-interviewer relationship. As discussed above, the particular social context of the interview — for example, the presence of a black interviewer — might cause respondents to moderate views they might consider “unacceptable” or could cause them to abstain from expressing any opinion altogether. While the results presented above would still be interesting if they were solely the result of the interaction between the race of the interviewer and the respondent, my conclusions concerning the broad reach of the social effects of the interview setting would be greatly tempered. Thus, in order to gauge whether the social effects I identify are simply a result of the specific social interaction with the interviewer, I ran a series of equations which included a measures of the race of the interviewer.24 The basic result of these models — a large ρ and an attenuated constant coefficient relative to the separate probit estimation — hold across the numerous model specifications. In an attempt to save the reader from the burden of another 30 pages of tables, I have omitted these results from the appendix. The results are, however, available from the author upon request. 23 A careful observation of the results in Tables 6 and 7 reveal that a number of other relationships change in significant ways once the correction for selection bias is implemented. For example, the predictive ability of the “value” variables — racial resentment, trust in government, and moral conservatism — are attenuated, in some cases by up to 20 percent. This result suggests that correcting for selection bias reduces the differences among groups in their attitudes concerning school integration. This makes the individuals in the sample appear to behave more alike than in the independent probit estimation — they all like school integration less. Similar trends appear to hold for other blocks of variables, such as the census region variables. The coefficient on the race variable, then, is unusual in that it is the only positively signed significant variable which increases in size across the two estimation techniques. 24 I initially attempted to use a series of dummy variables which accounted for both the race of the interviewer and the race of the respondent. However, because less than four percent of the respondents were questioned by black interviewers, these variables were highly collinear and could not be used in the analysis. I therefore settled on a simple “race-of-interviewer” dummy variable. This dummy variable captures all interviewer effects in the outcome equation because previous research demonstrates that both blacks (Anderson, Silver, and Abramson 1988, Davis 1997, Schuman and Converse 1971) and whites (Finkel, Guterbock, and Borg 1991, Kinder and Sanders 1996) give more racially liberal responses in the presence of black interviewers. I was, however, initially concerned that using the race-of-interviewer dummy variable would not properly capture the effects of the interaction between the 22 15 INSERT TABLES 8-10 ABOUT HERE As Tables 8-10 show, the introduction of this variable does not alter the basic conclusions identified above. While respondents appear to take a more liberal stand in the presence of a black interviewer, the race of the interviewer does not seem to systematically affect a respondent’s willingness to offer an opinion.25 Thus, though individuals may profess to hold stronger pro-integration stands when speaking to black interviewers, the introduction of the interviewer variable does not alter the basic conclusions regarding the presence and effects of selection bias identified above. Certainly, the effect of the race-of-interviewer variable on integration opinion is important in its own right. But the social desirability effects transcend the specific nature of the relationship between the interview and respondent. Many white respondents seem to hold their tongue rather than give what could be construed as a racist response, even when speaking to white interviewers. In short, the selection bias effects identified above are endemic to the survey research setting not to the particular respondent-interviewer interaction. Aggregate Consequences While the individual-level analyses presented above are interesting and important in their own right, ultimately, I am interested in discerning the effect of these individual-level processes on the shape of aggregate public opinion. Thus, to draw out the implication of my results for our understanding of collective public opinion on school integration I ran a series of simulations where I predicted aggregate support for government involvement in school integration. More specifically, I used my coefficient estimates from the equations reported above to generate each respondent’s predicted probability of supporting school integration. I then took the average of this probability across the sample to measure mean support for school integration.26 In order to gauge the aggregate effects of the selection bias in the question-answering process, I estimated the mean predicted probability of giving a supportive response under two conditions. I first estimated collective opinion distribution where selection effects are ignored (the independent probit interviewer and the respondent in the selection equation. This fear turned out to be baseless. A series of models in which I substituted a dummy variable scored “1” if both the interviewer and the respondent was white and “0” otherwise yielded identical results to the models where I used the race-of-interviewer variable (again, these results are omitted from the paper, but are available from the author upon request). 25 The 1990 data, it should be noted, do not replicate these results. In the models run using 1990 data, the presence of a black interviewer induce a respondent to be less likely to express an opinion, as my hypothesis predicts. But — interestingly enough — the presence of a black interviewer also lead respondents to take a more racially conservative position on the integration question. These results not only run counter to my results from 1992 and 1994, but run against previous research concerning race-of-interviewer effects. I cannot offer a definitive explanation for this result, but I suspect that his finding may simply be a fluke of the sample. This hypothesis is especially likely given that only 22 respondent in the sample used to estimate the model were interviewed by blacks. Thus, we should be careful about drawing general inferences from the “fluky” 1990 results. 26 To be precise, I ran my simulations in three steps. First, I used the estimates of the βs from the outcome equation of the separate probit model to generate a predicted value of Xi β for each individual in the sample. I then generated the predicted probability of supporting school integration, given the value of Xi β generated in the first step. Finally, I computed the mean opinion probability of giving a supportive response, across all individuals in the sample. I repeated this process for the estimates of the βs from the outcome equation of the bivariate probit selection model. 16 estimates). Second, I predicted opinion distribution where such effects are controlled for using the DubinRivers method (the bivariate probit estimates). In this way, the predicted behavior of the collective sample can be used to gauge the effect of controlling for selection bias.27 INSERT TABLES 11-13 ABOUT HERE As Table 11 demonstrates, correcting for selection bias greatly reduces projected support for school integration.28 No matter which model specification is chosen, the predicted support for school integration is much lower under the condition where selection bias is corrected relative to the “uncorrected” specification. In many cases, this difference is extremely significant. This trend also appears in the 1992 and 1990 data (see Table 12). In fact, accounting for selection bias yield even greater differences in those years. Similarly, as Table 13 demonstrates, introducing controls for the social context of the interview setting does not change the basic conclusions of the earlier results. In sum, then, the simulation results indicates that expressed opinion on the school integration question is a poor barometer of true support for integrationist policy. Rather, the fact that those hostile to integration gravitate to the “no opinion” responses means that expressed aggregate opinion in the NES surveys — which presents a near-even spilt on the question of support for government involvement in school integration in the 1990-1994 period — obscures strong public opposition to government action. An Aside (And An Independent Confirmation): The 1989 New York City Mayoral Race While my findings concerning the school integration question replicate well across different models and different data sets, the results are ultimately somewhat incomplete because there is no observable “check” on my results. That is, I can not compare the opposition to government involvement concerning school integration I find, using the selection bias model, to the observable behavior of the mass public. In this section, however, I demonstrate that the selection bias model I used above is also a valid predictor of true public sentiment on sensitive issues. Specifically, I show that my models of selection bias predict not only selection effects in the measurement of socially sensitive attitudes, but also selection bias in reports of individual behavior in socially sensitive situations. Just as opponents of government efforts to integrate schools seem to hide their opposition behind a veneer of indifference, pre-election polls in electoral contests which involve candidates of different races offer a situation where individuals might be loath to express their true candidate preferences for fear of appearing racist. Work by Reeves (1997) suggests that individuals who are uneasy or apprehensive about voting for black candidates may “vacate the field” in pre-election polls and declare themselves undecided I use for comparison the predicted values generated by the separate probit estimates, rather than the actual response frequency distributions, because I wish to control for the fact that my model of opinion formation is imperfect. By using predicted values generated by a model, I hold constant the predictive power of that model across the estimates of the aggregate public opinion. 28 I also ran a second simulation in which I estimated differences in the predicted behavior of respondents under the selection bias-corrected model relative to the separate equation model. Specifically I used the predicted probabilities of giving a supportive response to assign each respondent to the category which had the highest predicted probability of response. In other words, I assigned a respondent to the “support school integration” response if I predicted they has at least a 50 percent probability of supporting school integration, and assigned them to the “oppose school integration” category otherwise. I then computed mean opinion across these predicted responses. The results from this simulation were extremely similar to the predicted probability simulation presented above and, therefore, were omitted from the tables in the interests of parsimony. The simulation results are, however, available from the author upon request. 27 17 rather than come out and say that they oppose a black candidate. While Reeves does not demonstrate a direct link between the “undecided” option and support for white candidates, he presents the results of a survey experiment which suggests such a link. Specifically, he finds that the percentage of voters who claim to be “undecided” doubles in a hypothetical biracial contest compared to an identical race where both candidates are white. The fact that anti-black sentiment exerted a statistically significant effect on whites’ willingness to express their voting preferences in a bi-racial election suggests that an ”undecided” response may indeed be a covert vote for the white candidate. This effect, if true, would explain seemingly inexplicable pre-election polling results in two biracial contests races in the 1989 electoral season. In both the Virginia governor’s election and the New York City mayoral election, the black candidates (Wilder in Virginia and Dinkins in New York) held large leads over their white opponents (Coleman and Giuliani, respectively) in pre-election polls — which showed a large number of “undecided” voters — only to win their races by extremely narrow margins (see Finkel, Guterbock, and Borg 1991, Price and Traugott 1992, and Reeves 1997 for a discussion of these contests).29 In New York, for example, Dinkins held a 14-18 point advantage over Giuliani in polls taken only days before the election, but ended up winning the race by fewer than 50,000 votes out of the nearly two million ballots cast (Rosenthal 1989, McConnell 1990). The reason for the fallibility of the pre-election polls has been much debated (Clymer 1989, Rosenthal 1989). But, given the series of racially charged incidents in the summer of 1989 — lead by a racially motivated murder of a black youth in Bensonherst, Queens — it is clear that the political climate in New York made it difficult to remove the race factor from the electoral environment. This was especially true of certain groups — such as Jewish voters — who lay at the core of the Democratic party. Though there may have been many valid reasons to support Giuliani in the election, given the tenor of the times, it may have been difficult to perceive a Giuliani vote as anything but an anti-black vote (see McConnell 1990). Thus, many respondents may have had an incentive to say they were “undecided” rather than reveal their pro- Giuliani sentiments (for a post-election discussion of this possibility, see Clymer 1989 and Rosenthal 1989). The 1989 New York City mayoral race, then, parallels many of the conditions found in the context of the school integration question. In both cases, a large number of respondents abstained from offering their opinion or candidate choice. In both cases, the broader social climate made expressing a racially conservative opinion — a vote for Giuliani or opposition to government involvement in school integration — a socially difficult act. In the New York mayoral race, however, we have the advantage of being able to “check” how well the selection bias correction employed in the school integration data conforms to an observable outcome — in this case, the actual election results. Specifically, since we know that the preelection polls over-estimated Dinkins’s true level of support among the public we can see how well the predicted results from the selection bias correction conforms to the true election outcome. To gauge the presence and effect of selection bias in the 1989 mayoral pre-election polls, I used data from a poll conducted by ABC News and the New York Daily News on October 31 to November 6. This poll shows that 45.9 percent of registered voters support Dinkins, 35.4 percent support Giuliani, 4.8 percent support another candidate, and 14.0 percent are undecided. Thus, the pre-election poll indicates that Dinkins has 56.4 percent of the two-party vote — a seemingly commanding lead of almost 13 percentage points. These results conform to the results of other pre-elections polls (see Reeves 1997) but not to the true results of the election held just one day after the poll ended. It appears likely, then, that this polling data might be contaminated by the same type of selection bias found in the integration data. In the analyses below, I examine only data from the New York City mayoral race, because I have not yet found a data set from the Virginia election suitable for the selection bias model. 29 18 In order to determine whether, in fact, this hypothesis was true, I estimated two models of the candidate choice process using the 1989 mayoral data. In the first model (the independent probit) I used a simple probit to predict a respondent’s preference for Dinkins or Giuliani. In the second model (the bivariate probit model) I introduced the Dubin-Rivers correction for selection bias. Though I was somewhat limited in the choice of variables I could use in my model by the small size of the dataset, I found that the available variables were a more than adequate predictor of candidate choice. More importantly for the purposes of estimating the selection bias model, I also found a variable that identifies the selection equation. Specifically, the respondent’s self-reported certainty of voting in the mayoral race (controlling for previous voting behavior) was a strong predictor of whether one would offer a candidate choice, but did not affect the direction of that choice. Thus, my model posits that candidate choice (the outcome equation) is a function of the respondent’s: party of registration, ideology, education, age, religion, income, gender, and race. The willingness to offer a choice (the selection equation) is a function of these variables, as well as certainty of voting. INSERT TABLE 14 ABOUT HERE As Table 14 demonstrates, the bivariate probit selection model indicates that the 1989 New York pre-election polling data is indeed contaminated by the same selection bias as the school integration data. Rho is positive and highly significant in both a statistical and a substantive sense. This result indicates that those people who are likely to express a voting preference are — by simple dint of answering the preference question — more likely to say they support Dinkins. Conversely, the “don’t know” option conceals a large base of Giuliani support. Interestingly, however, unlike the school integration question, the effects of the selection bias in the pre-election polling data are not captured in the constant term. Rather, the selection bias affects the model coefficient estimates through the independent variables. For example, the uncorrected probit predicts that older respondents are more likely to vote for Giuliani than for Dinkins. But correcting for selection bias increases the predictive power of the age variable by over 50 percent. Similarly, the coefficient on the variable indicating whether the respondent is Jewish increases by over 20 percent once the selection bias correction is introduced.30 These results indicate that respondents who are older and Jewish — those groups who might be loath to express their anti-Dinkins sentiment — are more likely to support Giuliani then the uncorrected probit estimates suggests. Moreover, the coefficients on the “no ideology” and “refusal to reveal income” variables are greatly attenuated in the bivariate selection model results relative to the uncorrected estimates. Thus, people who are reluctant to reveal information about themselves are less likely to support Dinkins than would appear at first glance. None of the effects are surprising given the results of the selection equation. As Table 13 demonstrates, these groups of individuals — Jews, older respondents, and the reticent — are less likely than other individuals to volunteer their candidate choice to the survey interviewer. But while the results from the 1989 data may be different in type from the school integration results, they are similar in kind. In the case of school integration, the effect of the selection bias correction works through the constant term, indicating that all individuals in the sample are equally likely to hide their opposition to school integration behind a veneer of indifference.31 In the mayoral race, on the other hand, This result makes sense because these individuals are also less likely to reveal who they plan to vote for in the election. Accounting for the effects of their reticence on the coefficient estimates yields a pro-Giuliani stance that appears larger than it is at first glance. 31 Though, of course, the 1990 and 1992 NES results suggest that this generalization does not necessarily apply to black respondents. 30 19 the effect is larger for some groups of people — Jews, older respondents, and the reticent — than others. But in both cases, the result is the same; the “no-opinion” result seems to be a cover — for at least a significant proportion of the sample — for opposition to policies and candidates in choices that are racially sensitive. INSERT TABLE 15 ABOUT HERE This similarity between the school integration results and the 1989 pre-election poll results carries over to a simulation which predicts respondent’s candidate choice. As before, the average predicted probability of supporting Dinkins is estimated under two conditions — (1) the separate probit estimate and (2) the bivariate probit estimate. As Table 15 demonstrates, introducing the correction for selection bias greatly decreases the predicted support for Dinkins. In fact the point estimates of the predicted support for Dinkins under the bivariate probit results predict the actual electoral results almost perfectly. Thus, it seems that the selection bias correction compensates for the shortcomings of the pre-election polls. As noted above, the pre-election polls leading up to the 1989 New York City mayoral election predicted that Dinkins held a commanding lead over Giuliani, but the election returns proved that those polls were wrong. As Table 15 demonstrates, correcting the faulty polling data for selection bias yields results closer to the truth — that the race was, in fact, extremely tight. These analyses are especially important. Because the estimates corrected for selection bias do a better job of predicting the actual election results than the uncorrected estimates, it is clear that the selection bias correction provides a more reliable estimate of the true vote intentions than the preferences expressed in the survey. Given that the circumstances surrounding the mayoral election are similar to that of the school integration setup, we can then have greater confidence that the corrected estimates for the school integration data are closer to “the truth” as well. CONCLUSION Taken on their face, the results presented in this paper are illuminating. The unspoken opposition to government integration efforts revealed in my analyses indicates that latent mass opposition to school integration is higher than would appear from the marginals of opinion polls. This result may explain, in part, the unfinished legacy of the school integration efforts begun in the 1950s (Hochschild 1984). In addition, the results from the 1989 New York City mayoral analysis provide a compelling account of the failure of the pre-election polls in that race. Specifically, certain groups of individual —Jews, old respondents, and the reticent — chose to keep silent rather than voice their opposition to Dinkins. But while the analyses presented in this paper are interesting in their own right, these results also have broader implication for how we understand both the individual survey response and the collective signals sent by the mass public to eliets through opinion polls. At the individual level, my findings of selection bias in the school integration question-answering process demonstrates that — as the Mediated Communication model suggests — the opinions respondents express in the survey interview are not necessarily identical to the opinions they construct when coming to grips with a survey question. This result underscores the importance of attending to the effects of the social environment of the survey interview when investigating public opinion. The “new look” in public opinion research — led by the work of Zaller and Feldman, among others — has revolutionized our thinking about how individuals approach the cognitive tasks involved in the question-answering process. But it is now necessary to also revolutionize our understanding of the social side of the survey response and consider both the cognitive and social processes involved in the dual tasks of opinion formation and opinion expression. To a certain degree, the political science literature is beginning to address these concerns. The 20 growing literature concerning race-of-interviewer effects implicitly — if not explicitly — takes up the social concerns I have identified in this paper. But as the school integration results demonstrate, race-ofinterviewer effects do not tell the whole story. To fully address the complex effects arising from the social milieu of the survey interview, we must account for social processes in a comprehensive model of the survey response, such as the Mediated Communication model. The findings reported in this paper also have important implications for how we understand aggregate public opinion. The misrepresentation of individual opinion in the survey interview may be interesting in and of itself. But that misrepresentation acquires political bite to the extent that it biases the collective opinion signals which reach policymakers. The results reported in this paper suggest that, indeed, such bias exists in aggregate public opinion. Because the social environment of the survey interview systematically affects the willingness of respondents to answer survey questions on sensitive topics — here questions on racial attitudes — aggregate opinion polls may provide an inaccurate picture of true public sentiment on sensitive issues. This phenomenon is problematic for policy formation and implementation if eliets in any way use polls to serve as a barometer of public sentiment on sensitive policy controversies. At the same time, the analyses reported in this paper should not be taken as an indictment of the survey enterprise. While some individuals may misrepresent their political preferences in the environment of the survey interview, the message to take from this paper is not that we cannot gain valid information from polls. Instead, the results in these analyses highlight the importance of paying attention to and accounting for the types of biases which taint measures of aggregate public opinion. One important factor is the heterogeneity in the levels of social sensitivity that exists across issues. Thus, to poll effectively, we need to pay close attention to the content of an issue and consider how the social climate might affect the opinion distribution on that issue. But we can do more than think about this critical issue. In this paper, I provide a technique which allows us to not only speculate on the presence of social effects, but allows us to measure and — more importantly — to account for its bias. With a more critical and discerning eye towards the larger political and social environment, then, we may measure public opinion more effectively and move towards an understanding of the true shape of public sentiment on key issues facing the government. 21 TABLE 1: 1990-1994 SCHOOL INTEGRATION QUESTION OPINION DISTRIBUTION A: WITH DON’T KNOWS Year 1990 1992 1994 Support 31.9 31.9 27.4 Oppose 30.4 32.7 35.1 Don’t Know 37.7 35.3 37.5 B: DON’T KNOWS EXCLUDED Year 1990 1992 1994 Support 51.2 49.4 43.8 Oppose 48.8 50.6 56.2 Source: National Election Study Cumulative File 22 TABLE 2: 1994 SCHOOL INTEGRATION QUESTION Selection: Engagement Model Outcome: Demographic Model Variable Independent Probit Coefficient (SE) OUTCOME EQUATION -0.074 (0.264) -0.006 (0.003)** 1.179 (0.157)** 0.474 (0.166)** -0.314 (0.094)** -0.086 (0.101) -0.026 (0.212) 0.292 (0.177)* 0.176 (0.189) 0.342 (0.148)** -0.051 (0.145) -0.149 (0.152) 0.183 (0.152) 0.105 (0.193) -0.057 (0.137) 0.117 (0.142) 0.125 (0.149) -0.042 (0.134) 0.398 (0.133)** 0.204 (0.112)* 0.488 (0.276)** 0.023 (0.134) 0.132 (0.159) -0.119 (0.179) 0.016 (0.166) -0.584 (0.227)** 0.209 (0.193) -0.071 (0.180) 0.054 (0.162) SELECTION EQUATION — — — — — — — — — CORRELATION PARAMETERS — 1029/-620.863 Bivariate Probit Coefficient (SE) -0.639 (0.224)** -0.004 (0.003) 1.020 (0.175)** 0.369 (0.136)** -0.228 (0.082)** -0.059 (0.080) 0.114 (0.173) 0.204 (0.145) 0.118 (0.151) 0.267 (0.123)** -0.022 (0.119) -0.109 (0.126) 0.167 (0.142) 0.052 (0.160) -0.054 (0.111) 0.114 (0.117) 0.083 (0.120) -0.023 (0.111) 0.317 (0.112)** 0.150 (0.096) 0.438 (0.228)** -0.002 (0.107) 0.118 (0.132) -0.099 (0.145) -0.020 (0.130) -0.484 (0.199)** 0.164 (0.160) -0.040 (0.150) 0.015 (0.134) -0.075 (0.120) 0.391 (0.132)** 0.408 (0.099)** -0.190 (0.079)** 0.234 (0.106)** 0.157 (0.063)** 0.158 (0.112) -0.168 (0.141) 0.028 (0.056) 0.812 (0.159)** 1647/-1673.928 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Constant Political Information Discuss Politics No Ideology Black Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls ρ N/Log Likelihood * = p < .10; ** = p < .05 23 TABLE 3: 1994 SCHOOL INTEGRATION QUESTION Selection: Full Model Outcome: Demographic Model Variable Independent Probit Coefficient (SE) OUTCOME EQUATION -0.074 (0.264) -0.006 (0.003)* 1.179 (0.157)** 0.474 (0.166)** -0.314 (0.094)** -0.086 (0.101) -0.026 (0.212) 0.292 (0.177)* 0.176 (0.189) 0.342 (0.148)** -0.051 (0.145) -0.149 (0.152) 0.105 (0.192) 0.105 (0.193) -0.057 (0.137) 0.117 (0.143) 0.125 (0.149) -0.042 (0.134) 0.397 (0.133)** 0.204 (0.112)* 0.488 (0.276)* 0.023 (0.134) 0.132 (0.159) -0.119 (0.179) 0.155 (0.166) -0.584 (0.227)** 0.208 (0.193) -0.071 (0.180) 0.054 (0.162) Bivariate Probit Coefficient (SE) -0.703 (0.244)** -0.002 (0.003) 1.028 (0.183)** 0.370 (0.148)** -0.247 (0.088)** -0.107 (0.087) 0.249 (0.191) 0.148 (0.163) 0.027 (0.166) 0.213 (0.139) -0.076 (0.128) -0.166 (0.135) 0.115 (0.157) 0.105 (0.193) -0.057 (0.121) 0.201 (0.126) 0.085 (0.129) -0.021 (0.111) 0.308 (0.122)** 0.226 (0.100)** 0.519 (0.251)** 0.057 (0.116) 0.030 (0.146) -0.158 (0.158) 0.148 (0.145) -0.604 (0.203)** 0.145 (0.173) -0.091 (0.163) -0.101 (0.147) Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other 24 TABLE 3 (CONTINUED): Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 1029/-620.863 Bivariate Probit Coefficient (SE) -0.265 (0.229) 0.006 (0.003)** 0.262 (0.117)** -0.017 (0.136) -0.013 (0.076) -0.097 (0.080) 0.422 (0.182)** -0.139 (0.140) -0.224 (0.143)* -0.123 (0.120) -0.119 (0.116) -0.136 (0.122) -0.112 (0.145) -0.161 (0.147) -0.019 (0.107) 0.215 (0.114)* 0.006 (0.101) 0.006 (0.104) 0.174 (0.086)** 0.194 (0.245) 0.143 (0.103) -0.183 (0.126) -0.059 (0.133) -0.220 (0.113)** -0.293 (0.157)** -0.034 (0.154) -0.092 (0.145) -0.240 (0.124)* 0.207 (0.155) 0.387 (0.101)** -0.155 (0.082)** 0.169 (0.067)** 0.143 (0.116) -0.165 (0.145) 0.030 (0.059) 0.826 (0.177)** 1647/-1041.177 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls ρ N/Log Likelihood * = p < .10; ** = p < .05 25 TABLE 4: 1994 SCHOOL INTEGRATION QUESTION Selection: Full Model Outcome: Combined Model Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 0.549 (0.438) -0.009 (0.004)** 0.633 (0.175)** 0.374 (0.172)** -0.145 (0.102) 0.050 (0.108) -0.353 (0.244) 0.255 (0.187) 0.025 (0.199) 0.264 (0.158)* -0.055 (0.155) -0.036 (0.163) 0.278 (0.188) 0.111 (0.204) -0.011 (0.148) 0.189 (0.154) 0.195 (0.163) -0.022 (0.145) 0.186 (0.158) 0.069 (0.122) 0.200 (0.306) -0.049 (0.142) 0.233 (0.169) -0.002 (0.189) 0.017 (0.175) -0.593 (0.257)** 0.318 (0.203) 0.099 (0.192) 0.142 (0.172) -0.197 (0.097)** -1.121 (0.236)** 0.961 (0.288)** 0.510 (0.215)** -0.649 (0.256)** 0.218 (0.139) 0.217 (0.082)** -0.277 (0.148)* -0.249 (0.121)** -0.079 (0.134) Bivariate Probit Coefficient (SE) 0.093 (0.586) -0.006 (0.005) 0.653 (0.177)** 0.360 (0.164)** -0.131 (0.101) 0.016 (0.109) -0.189 (0.288) 0.189 (0.193) -0.044 (0.203) 0.209 (0.166)* -0.078 (0.151) -0.068 (0.158) 0.238 (0.189) 0.072 (0.205) -0.014 (0.147) 0.242 (0.157) 0.195 (0.163) -0.014 (0.135) 0.172 (0.157) 0.103 (0.121) 0.255 (0.310) -0.009 (0.140) 0.167 (0.179) -0.040 (0.181) 0.082 (0.193) -0.636 (0.258)** 0.287 (0.199) 0.068 (0.191) 0.049 (0.195) -0.146 (0.113) -1.039 (0.265)** 0.883 (0.304)** 0.460 (0.223)** -0.605 (0.254)** 0.208 (0.133) 0.199 (0.084)** -0.246 (0.146)* -0.223 (0.126)* -0.145 (0.143) Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology 26 TABLE 4 (CONTINUED): Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 1009/-542.516 Bivariate Probit Coefficient (SE) -0.244 (0.238) 0.007 (0.003)** 0.264 (0.119)** -0.018 (0.140) 0.012 (0.078) -0.098 (0.083) 0.383 (0.186)** -0.128 (0.144) -0.240 (0.146)* -0.146 (0.123) -0.110 (0.118) -0.135 (0.124) -0.096 (0.149) -0.082 (0.155) 0.005 (0.109) 0.235 (0.116)** 0.014 (0.103) -0.027 (0.107) 0.167 (0.087)** 0.189 (0.243) 0.145 (0.107) -0.166 (0.128) -0.141 (0.146) -0.315 (0.127)** -0.320 (0.159) -0.022 (0.157) -0.094 (0.148) -0.285 (0.126)** 0.133 (0.165) 0.403 (0.108)** -0.173 (0.090)** 0.127 (0.074)** 0.131 (0.123) -0.177 (0.155) 0.040 (0.065) 0.518 (0.458) 1600/-1548.610 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls ρ N/Log Likelihood * = p < .10; ** = p < .05 27 TABLE 5: 1994 SCHOOL INTEGRATION QUESTION A: VALUES OF RHO ACROSS MODEL SPECIFICATIONS SELECTION EQUATION Values Engagement Full 0.828 (0.154)** 0.670 (0.203)** OUTCOME EQUATION Demographic 0.812 (0.159)** 0.827 (0.177)** Self-Interest 0.826 (0.177)** 0.839 (0.141)** Combined 0.631 (0.326)** 0.518 (0.457) B: DIFFERENCE IN CONSTANT ACROSS MODEL SPECIFICATIONS (SEPARATE PROBIT-BIVARIATE PROBIT) SELECTION EQUATION Values Engagement Full .413 .381 OUTCOME EQUATION Demographic .565 .629 Self-Interest .406 .346 Combined .492 .456 Note: Standard Errors in Parenthesis * = p < .10; ** = p < .05 28 TABLE 6: 1992 SCHOOL INTEGRATION QUESTION Selection: Full Model Outcome: Combined Model Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 0.620 (0.442) 0.003 (0.003) 0.272 (0.146)* 0.849 (0.143)** -0.186 (0.092)** -0.019 (0.095) 0.077 (0.204) 0.036 (0.160) 0.015 (0.158) -0.107 (0.147) -0.257 (0.144)* -0.270 (0.148)* -0.162 (0.178) -0.228 (0.199) -0.212 (0.121)* 0.024 (0.136) -0.275 (0.138)** 0.174 (0.129) 0.065 (0.135) 0.052 (0.135) -0.018 (0.278) -0.057 (0.143) -0.181 (0.155) -0.166 (0.191) -0.081 (0.158) -0.171 (0.180) 0.279 (0.182) -0.109 (0.174) -0.182 (0.145) -0.590 (0.362)* -2.120 (0.378)** 1.370 (0.256)** 1.029 (0.194)** -0.800 (0.222)** 0.183 (0.124) 0.167 (0.073)** -0.119 (0.123) 0.084 (0.109) 0.066 (0.120) Bivariate Probit Coefficient (SE) 0.048 (0.433) 0.004 (0.003) 0.332 (0.130)** 0.746 (0.2096)** -0.132 (0.089) -0.004 (0.088) -0.008 (0.184) 0.067 (0.147) 0.074 (0.143) -0.113 (0.131) -0.250 (0.131)* -0.264 (0.134)** -0.103 (0.183) -0.303 (0.174)* -0.158 (0.114) 0.042 (0.119) -0.187 (0.130)* 0.150 (0.118) 0.013 (0.124) 0.041 (0.098) -0.095 (0.266) -0.066 (0.131) -0.088 (0.148) -0.137 (0.183) -0.062 (0.143) -0.133 (0.164) 0.294 (0.170)* -0.078 (0.161) -0.185 (0.132) -0.538 (0.341) -1.769 (0.403)** 1.130 (0.262)** 0.846 (0.202)** -0.658 (0.207)** 0.158 (0.110) 0.145 (0.064)** -0.093 (0.106) 0.082 (0.094) -0.106 (0.122) Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology 29 TABLE 6 (CONTINUED): Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 1255/-683.571 Bivariate Probit Coefficient (SE) 0.155 (0.213) 0.003 (0.002) 0.325 (0.104)** 0.146 (0.136) 0.000 (0.072) 0.035 (0.072) -0.345 (0.165)** 0.102 (0.123) 0.136 (0.124) -0.067 (0.111) -0.124 (0.109) -0.117 (0.112) 0.022 (0.143) -0.271 (0.144)* 0.018 (0.093) 0.058 (0.103) 0.085 (0.092) -0.107 (0.096) 0.005 (0.082) 0.179 (0.234) -0.079 (0.107) 0.157 (0.119) 0.013 (0.137) 0.054 (0.118) 0.072 (0.136) 0.176 (0.141) 0.043 (0.130) -0.052 (0.109) 0.455 (0.151)** 0.300 (0.100)** -0.312 (0.078)** -0.143 (0.278) 0.101 (0.253) -0.004 (0.128) -0.003 (0.060) 0.758 (0.218)** 1915/-1875.684 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls ρ .N/Log Likelihood * = p < .10; ** = p < .05 30 TABLE 7: 1990 SCHOOL INTEGRATION QUESTION Selection: Full Model Outcome: Combined Model Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 1.064 (0.619)* -0.008 (0.005) 0.305 (0.217) 0.374 (0.264) 0.095 (0.141) -0.218 (0.148) 0.051 (0.316) 0.028 (0.263) -0.204 (0.252) 0.038 (0.221) -0.049 (0.209) 0.033 (0.247) -0.170 (0.278) 0.399 (0.278) -0.464 (0.194)** -0.068 (0.199) 0.052 (0.220) -0.033 (0.200) -0.039 (0.154) 0.110 (0.159) -0.190 (0.494) 0.411 (0.213)* -0.147 (0.230) -0.419 (0.303) -0.122 (0.228) -0.039 (0.268) -0.520 (0.266)** 0.087 (0.262) 0.130 (0.244) 0.254 (0.141)* -1.462 (0.300)** 0.719 (0.387)* 0.746 (0.286)** -0.768 (0.363)** 0.176 (0.195) -0.004 (0.101) 0.035 (0.193) -0.371 (0.171)** 0.101 (0.179) Bivariate Probit Coefficient (SE) 0.284 (0.795) -0.005 (0.005) 0.480 (0.213)** 0.307 (0.243) 0.083 (0.134) -0.200 (0.141) 0.201 (0.321) -0.079 (0.243) -0.275 (0.231) -0.135 (0.228) -0.119 (0.209) -0.052 (0.255) -0.245 (0.263) 0.278 (0.274) -0.366 (0.208)* -0.005 (0.194) 0.164 (0.224) -0.048 (0.189) 0.049 (0.165) 0.152 (0.154) -0.206 (0.562) 0.338 (0.236) 0.006 (0.254) -0.385 (0.311) -0.053 (0.218) -0.111 (0.283) -0.348 (0.286) 0.112 (0.248) 0.145 (0.244) 0.255 (0.137)* -1.274 (0.355)** 0.642 (0.379)* 0.648 (0.295)** -0.595 (0.379) 0.146 (0.183) 0.006 (0.094) 0.035 (0.172) -0.327 (0.164)** -0.069 (0.201) Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology 31 TABLE 7 (CONTINUED) Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 555/-293.599 Bivariate Probit Coefficient (SE) -0.165 (0.332) 0.004 (0.003) 0.661 (0.162)** -0.039 (0.107) -0.039 (0.107) -0.061 (0.114) 0.229 (0.245) -0.267 (0.193) -0.245 (0.200) -0.444 (0.179)** -0.262 (0.178) -0.269 (0.209) -0.177 (0.232) -0.177 (0.232) 0.075 (0.143) 0.137 (0.153) 0.271 (0.144)* 0.207 (0.109)* 0.163 (0.120) -0.289 (0.437) -0.071 (0.147) 0.327 (0.179)* -0.141 (0.208) 0.140 (0.166) 0.409 (0.213)* 0.231 (0.215) -0.036 (0.190) -0.002 (0.167) 0.505 (0.329) 0.417 (0.166)** -0.398 (0.108)** 0.091 (0.107) 0.426 (0.250)* 0.200 (0.130) -0.060 (0.087) 0.677 (0.380)* 889/-832.323 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls ρ .N/Log Likelihood * = p < .10; ** = p < .05 32 TABLE 8: 1994 SCHOOL INTEGRATION QUESTION Selection: Full Model + Race-of-interviewer Variable Outcome: Combined Model + Race-of-interviewer Variable Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 0.495 (0.440) -0.008 (0.004)** 0.570 (0.177)* 0.385 (0.172)** -0.145 (0.102) 0.026 (0.102) -0.343 (0.245) 0.248 (0.187) 0.010 (0.199) 0.250 (0.159) -0.055 (0.155) -0.032 (0.164) 0.290 (0.188) 0.074 (0.206) 0.022 (0.150) 0.228 (0.156) 0.237 (0.165) -0.027 (0.145) 0.186 (0.158) 0.066 (0.122) 0.170 (0.305) -0.056 (0.143) 0.229 (0.169) 0.003 (0.189) -0.008 (0.177) -0.653 (0.263)** 0.319 (0.203) 0.108 (0.192) 0.151 (0.172) -0.188 (0.097)** -1.072 (0.237)** 0.958 (0.289)** 0.487 (0.216)** -0.665 (0.257)** 0.205 (0.139) 0.215 (0.083)** -0.276 (0.148)* -0.248 (0.121)** -0.075 (0.140) 0.786 (0.355)** Bivariate Probit Coefficient (SE) 0.062 (0.604) -0.006 (0.005) 0.588 (0.181)** 0.370 (0.165)** -0.131 (0.102) 0.000 (0.110) -0.190 (0.295) 0.187 (0.194) -0.056 (0.207) 0.198 (0.166) -0.076 (0.153) -0.063 (0.161) 0.253 (0.191) 0.039 (0.206) 0.018 (0.151) 0.279 (0.161)* 0.220 (0.167) -0.019 (0.137) 0.173 (0.158) 0.098 (0.122) 0.223 (0.305) -0.018 (0.142) 0.167 (0.182) 0.032 (0.182) -0.103 (0.195) -0.692 (0.261)** 0.291 (0.200) 0.079 (0.193) 0.062 (0.198) -0.139 (0.113) -1.000 (0.263)** 0.886 (0.307)** 0.442 (0.224)** -0.624 (0.258)** 0.199 (0.134) 0.199 (0.085)** -0.247 (0.150) -0.225 (0.127)* -0.139 (0.146) 0.782 (0.368)** Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology Interviewer Black 33 TABLE 8 (CONTINUED): 1994 SCHOOL INTEGRATION QUESTION Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 1009/-539.723 Bivariate Probit Coefficient (SE) -0.247 (0.239) 0.007 (0.003)** 0.236 (0.124)** 0.014 (0.140) 0.013 (0.078) -0.102 (0.083) 0.385 (0.186)** -0.127 (0.144) -0.245 (0.146)* -0.152 (0.124) -0.111 (0.118) -0.136 (0.124) -0.094 (0.149) -0.090 (0.155) 0.012 (0.109) 0.245 (0.117)** 0.024 (0.104) -0.027 (0.107) 0.168 (0.087)** 0.186 (0.243) 0.143 (0.107) -0.170 (0.128) -0.141 (0.146) -0.320 (0.128)** -0.323 (0.160)** -0.022 (0.158) -0.094 (0.148) -0.284 (0.126)** 0.126 (0.166) 0.404 (0.108)** -0.173 (0.090)* 0.129 (0.075)* 0.132 (0.123) -0.174 (0.156) 0.038 (0.066) 0.218 (0.228) 0.497 (0.481) 1600/-1545.415 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls Interviewer Black ρ N/Log Likelihood * = p < .10; ** = p < .05 34 TABLE 9: 1992 SCHOOL INTEGRATION QUESTION Selection: Full Model + Race-of-interviewer Variable Outcome: Combined Model + Race-of-interviewer Variable Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 0.509 (0.445) 0.003 (0.003) 0.185 (0.150) 0.847 (0.195)** -0.176 (0.092)** -0.033 (0.096) 0.086 (0.206) 0.048 (0.161) 0.049 (0.159) -0.099 (0.148) -0.227 (0.146) -0.232 (0.148) -0.139 (0.179) -0.195 (0.200) -0.203 (0.122)* 0.062 (0.134) -0.222 (0.139) 0.150 (0.130) 0.074 (0.135) 0.060 (0.109) -0.023 (0.280) -0.093 (0.144) -0.175 (0.155) -0.162 (0.191) -0.073 (0.160) -0.181 (0.181) 0.278 (0.184) -0.111 (0.175) -0.183 (0.146) -0.573 (0.364) -2.084 (0.379)** 1.400 (0.256)** 1.054 (0.194)** -0.804 (0.223)** 0.181 (0.124) 0.150 (0.073)** -0.145 (0.124) 0.075 (0.109) 0.073 (0.120) 1.029 (0.338)** Bivariate Probit Coefficient (SE) -0.044 (0.432) 0.004 (0.003) 0.265 (0.136)** 0.747 (0.207)** -0.124 (0.090) -0.007 (0.088) -0.003 (0.186) 0.077 (0.148) 0.102 (0.144) -0.105 (0.133) -0.226 (0.131)* -0.231 (0.131)* -0.084 (0.169) -0.274 (0.186) -0.150 (0.114) 0.074 (0.118) -0.144 (0.129) 0.131 (0.120) 0.022 (0.134) 0.050 (0.099) 0.086 (0.267) -0.099 (0.134) -0.086 (0.149) -0.134 (0.184) -0.057 (0.145) -0.142 (0.165) 0.287 (0.172)* -0.080 (0.163) -0.186 (0.133) -0.527 (0.343) -1.742 (0.407)** 1.163 (0.265)** 0.870 (0.202)** -0.664 (0.209)** 0.157 (0.110) 0.131 (0.064)** -0.115 (0.108) 0.074 (0.095) -0.102 (0.124) 0.772 (0.344)** Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology Interviewer Black 35 TABLE 9 (CONTINUED): Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 1255/-677.848 Bivariate Probit Coefficient (SE) 0.177 (0.200) 0.003 (0.003) 0.340 (0.107)** 0.152 (0.138) -0.001 (0.072) 0.036 (0.072) 0.349 (0.164)** 0.103 (0.123) 0.132 (0.125) -0.070 (0.111) -0.128 (0.109) -0.121 (0.112) 0.018 (0.143) -0.274 (0.145)* 0.016 (0.093) 0.051 (0.103) 0.078 (0.093) -0.107 (0.097) 0.004 (0.082) 0.182 (0.234) -0.077 (0.108) 0.160 (0.119) 0.013 (0.137) 0.054 (0.118) 0.074 (0.136) 0.178 (0.141) 0.039 (0.130) 0.056 (0.109) 0.462 (0.152)** 0.296 (0.100)** -0.314 (0.079)** -0.150 (0.278) 0.108 (0.256) -0.012 (0.128) -0.008 (0.045) -0.103 (0.185) 0.753 (0.224)** 1915/-1869.802 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls Interviewer Black ρ N/Log Likelihood * = p < .10; ** = p < .05 36 TABLE 10: 1990 SCHOOL INTEGRATION QUESTION Selection: Full Model + Race-of-interviewer Variable Outcome: Combined Model + Race-of-interviewer Variable Variable Independent Probit Coefficient (SE) OUTCOME EQUATION 1.027 (0.622)* -0.008 (0.005) 0.395 (0.227)* 0.430 (0.267)* 0.081 (0.141) -0.195 (0.150) 0.065 (0.318) 0.066 (0.264) -0.196 (0.252) 0.066 (0.222) -0.027 (0.210) 0.025 (0.248) -0.160 (0.278) 0.439 (0.280) -0.445 (0.195)** -0.089 (0.200) 0.053 (0.220) -0.053 (0.200) -0.024 (0.154) 0.112 (0.159) -0.217 (0.492) 0.417 (0.214)** -0.133 (0.231) -0.420 (0.303) -0.120 (0.229) -0.033 (0.270) -0.482 (0.269)* 0.079 (0.263) 0.174 (0.247) 0.251 (0.143)* -1.479 (0.301)** 0.718 (0.389)* 0.731 (0.287)** -0.743 (0.364)** 0.163 (0.196) -0.004 (0.101) 0.068 (0.195) -0.361 (0.172)** 0.083 (0.180) -0.579 (0.400) Bivariate Probit Coefficient (SE) 0.286 (0.811) -0.005 (0.005) 0.566 (0.225)** 0.366 (0.254) 0.067 (0.137) -0.180 (0.143) 0.212 (0.328) -0.036 (0.251) -0.264 (0.234) -0.102 (0.233) -0.096 (0.212) -0.053 (0.256) -0.231 (0.265) 0.328 (0.287) -0.352 (0.210)* -0.030 (0.199) 0.162 (0.228) -0.052 (0.191) 0.061 (0.167) 0.151 (0.156) -0.235 (0.560) 0.347 (0.243) 0.013 (0.258) -0.393 (0.314) -0.054 (0.224) -0.109 (0.286) -0.309 (0.294) 0.100 (0.251) 0.188 (0.227) 0.254 (0.139)* -1.304 (0.363)** 0.647 (0.388)* 0.642 (0.300)** -0.582 (0.381) 0.136 (0.185) 0.005 (0.010) 0.064 (0.179) -0.321 (0.166)** -0.082 (0.204) -0.678 (0.352)** Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South Grew Up in South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Number of School-Age Children Racial Resentment Equality Trust in Government Moral Conservatism Religious Importance Party Identification Liberal Conservative No Ideology Interviewer Black 37 TABLE 10 (CONTINUED): Variable Independent Probit Coefficient (SE) SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 555/-292.547 Bivariate Probit Coefficient (SE) -0.180 (0.333) 0.004 (0.003) 0.701 (0.161)** 0.007 (0.166) -0.054 (0.108) -0.054 (0.115) 0.237 (0.247) -0.252 (0.194) -0.238 (0.201) -0.437 (0.180)** -0.260 (0.179) -0.270 (0.211) -0.271 (0.230) -0.149 (0.235) 0.083 (0.143) 0.126 (0.154) 0.277 (0.144)* 0.214 (0.109)** 0.159 (0.121) -0.303 (0.439) -0.079 (0.147) 0.334 (0.179)* -0.146 (0.208) 0.141 (0.166) 0.405 (0.213)* 0.290 (0.225) -0.049 (0.190) 0.007 (0.167) 0.532 (0.335) 0.417 (0.168)** -0.412 (0.109)** 0.094 (0.108) 0.425 (0.249)* 0.210 (0.131) -0.058 (0.088) -0.513 (0.333) 0.650 (0.401)* 889/-829.866 Constant Age Black Hispanic Male Homeowner Education Income: <$10,000 Income: $10,000-$14,999 Income: $15,000-$24,999 Income: $35,000-$49,999 Income: $50,000-$74,999 Income: $75,000+ Income Not Ascertained North-Central West South No Religion Catholic Jewish Other Religion Occupation: Professional Occupation: Manager Occupation: White-Collar Occupation: Self Employed Occupation: Skilled Worker Occupation: Homemaker Occupation: Other Political Information Discuss Politics No Ideology Number of School-Age Children Refusal Conversion Persuasion Letter Number of Calls Interviewer Black ρ N/Log Likelihood * = p < .10; ** = p < .05 38 TABLE 11: PREDICTED OUTCOMES 1994 SCHOOL INTEGRATION QUESTION Specification (selection/outcome) Engagement/Values Engagement/Demographic Engagement/Self-Interest Engagement/Combined Full/Values Full/Demographic Full/Self-Interest Full/Combined Independent Probit 0.441 0.448 0.443 0.401 0.447 0.448 0.443 0.401 Bivariate Probit 0.296 0.305 0.289 0.330 0.318 0.306 0.291 0.356 Difference 0.145 0.143 0.154 0.071 0.129 0.142 0.152 0.045 39 TABLE 12: PREDICTED PROBABILITIES 1990-1992 SCHOOL INTEGRATION QUESTION Specification (selection/outcome) 1992: Full/Combined 1990: Full/Combined Independent Probit 0.494 0.582 Bivariate Probit 0.357 0.375 Difference 0.137 0.207 TABLE 13: PREDICTED PROBABILITIES 1990-1994 SCHOOL INTEGRATION QUESTION WITH RACE-OF-INTERVIEWER VARIABLE Specification (selection/outcome) 1994: Full/Combined (+ROI) 1992: Full/Combined (+ROI) 1990: Full/Combined (+ROI) Independent Probit 0.425 0.494 0.584 Bivariate Probit 0.355 0.359 0.377 Difference 0.070 0.135 0.207 40 TABLE 14: 1989 NEW YORK MAYORAL RACE: VOTE CHOICE Variable Independent Probit Coefficient (SE) OUTCOME EQUATION -0.066 (0.204) -1.146 (0.096)** -0.499 (0.116)** -0.215 (0.234) 0.449 (0.080)** -0.341 (0.088)** 0.542 (0.241)** 0.467 (0.156)** -0.004 (0.002)* -0.412 (0.101)** -0.277 (0.114)** -0.121 (0.144) -0.443 (0.281) 0.366 (0.168)** 0.494 (0.167)** 0.196 (0.132) 0.164 (0.106) -0.063 (0.092) 0.325 (0.144)** 0.005 (0.069) 2.110 (0.140)** 0.791 (0.112)** 1.458 (0.236)** Bivariate Probit Coefficient (SE) -0.099 (0.201) -1.040 (0.109)** -0.467 (0.115)** -0.200 (0.235) 0.464 (0.078)** -0.310 (0.091)** 0.285 (0.255) 0.444 (0.152)** -0.006 (0.002)** -0.390 (0.103)** -0.326 (0.115)** -0.143 (0.138) -0.502 (0.272)* 0.313 (0.169)* 0.437 (0.178)** 0.173 (0.130) 0.152 (0.105) -0.036 (0.089) 0.152 (0.182) -0.022 (0.069) 2.046 (0.203)** 0.807 (0.108)** 1.424 (0.243)** 1.416 (0.249)** 0.405 (0.405)** 0.035 (0.131) 0.050 (0.242) 0.214 (0.089)** 0.045 (0.093) -0.544 (0.159)** 0.073 (0.167) -0.010 (0.002)** 0.006 (0.119) -0.310 (0.126)** -0.101 (0.136) -0.246 (0.228) -0.130 (0.157) -0.132 (0.166) -0.087 (0.138) -0.033 (0.113) 0.102 (0.106) -0.552 (0.121)** -0.130 (0.074)* 0.482 (0.123)** 0.394 (0.162)** 0.287 (0.249) 0.332 (0.165)** -0.122 (0.117) 0.658 (0.302)** 2509/-1748.439 Constant Republican Independent Other Party Identifier Liberal Conservative No Ideology Education Age Catholic Jewish Other Religion Refused to Identify Religion Income: <$8,000 Income: $8,000-$11,999 Income: $12,000-$19,999 Income: $20,000-$29,999 Income: $50,000+ Income Not Ascertained Female Black Hispanic (White and Other) Black Hispanic Constant Republican Independent Other Party Identifier Liberal Conservative No Ideology Education Age Catholic Jewish Other Religion Refused to Identify Religion Income: <$8,000 Income: $8,000-$11,999 Income: $12,000-$19,999 Income: $20,000-$29,999 Income: $50,000+ Income Not Ascertained Female Black Hispanic (Not Black Hispanic) Black Hispanic Certainty of Voting Voted in 1988 ρ N/Log Likelihood SELECTION EQUATION — — — — — — — — — — — — — — — — — — — — — — — — — CORRELATION PARAMETERS — 2192/-911.708 * = p < .10; ** = p < .05 41 TABLE 15: PREDICTED PROBABILITIES 1989 NEW YORK CITY MAYORAL RACE: VOTE FOR DINKINS Comparison Predicted Probability of A Supportive Response Independent Probit 0.553 Bivariate Probit 0.512 Difference 0.041 42 APPENDIX A: CODING PROTOCOL, NES VARIABLES Age Black Age of the respondent, in years. Dummy indicating the race of the respondent (0=nonblack; 1=black) Dummy indicating whether the respondent is Hispanic (0=non-Hispanic; 1=Hispanic) Dummy indicating the gender of the respondent (0=female; 1=male) Dummy indicating whether the respondent owns their home (0=no; 1=yes) 7 category NES education variable measuring highest level of education (0=grade school; 1=advanced degree). A series of 8 dummy variables indicating the respondents reported income, or if the respondent’s income was not ascertained. The omitted category is an income of $25,000-$34,999 A Series of 4 dummy variables indicating the respondent’s census region of residence (North-East, North-Central, West, and South). The omitted category is respondents who live in the North-East Dummy indicating whether the respondent grew up in the South (0=no; 1=yes) A series of 5 dummy variables indicating the respondent’s religion (Protestant, Catholic, Jewish, other religion, and no religion). The omitted category is respondents who are Protestants. A series of 8 dummy variables indicating the respondent’s occupation, according to classifications developed by Hout et al (1995) (professional, manager, white-collar, self-employed, skilled worker, un- or semi-skilled homemaker, and other). The omitted category is respondents who are un- or semi-skilled. Hispanic Male Homeowner Education Income Variables Region Variables Grew Up in South Religion Variables Occupation Variables 43 APPENDIX A: CODING PROTOCOL, NES VARIABLES (CONTINUED) Number of School Age Children The natural log of the number of children in the household ages 6-17 (+1). 5 category partisanship variable. This variable is simply the NES 7-Category partisanship variable with the independent leaners collapsed with the weak partisans (-1=Strong Republican; 1=Strong Democrat). Dummy indicating self-identification as “extremely liberal,” “liberal,” or “slightly liberal” on the NES sevenpoint ideology scale. (0=not liberal; 1=liberal) Dummy indicating self-identification as “extremely conservative,” “conservative,” or “slightly conservative” on the NES seven-point ideology scale. (0=not conservative; 1=conservative) Dummy indicating self-identification as “moderate; middle of the road” on the NES seven-point ideology scale. (0=Not moderate; 1=moderate) Dummy indicating respondents who “don’t know” or “haven’t thought much about” where they place on the NES seven-point ideology scale. (0=claim ideology; 1=claim no ideology) 6 category NES Equality scale. Respondents are assigned their mean score across all the individual equality items as long as they answer half or more of those items. (0=low support for equality; 1=high support for equality) 4 category NES Trust in Government scale. Respondents are assigned their mean score across all the individual limited government items as long as they answer half or more of those items. (0=low trust in government; 1=high trust in government) 4 category NES Moral Conservatism scale. Respondents are assigned their mean score across all the individual moral conservatism items as long as they answer half or more of those items. (0=low support for moral conservatism; 1=high support for moral conservatism) Party Identification Liberal Conservative Moderate No Ideology Equality Trust in Government Moral Conservatism 44 APPENDIX A: CODING PROTOCOL, NES VARIABLES (CONTINUED) Religious Importance 4 category variable which gauges the degree of guidance religion provides in the respondent’s everyday life. (0=Not important; 1=provides a great deal of guidance) 9 category NES variable measuring knowledge of politics. (0=low; 1=high). 8 category variable measuring the number of days in the past week in which the respondent discussed politics with their friends and family, recoded to the 0-1 interval. (0=never discuss politics; 1=discuss politics every day). The natural log of the number of face-to-face and telephone calls made to the respondents home in order to obtain the interview. Dummy variable indicating whether the interviewer attempted to convert a respondent who initially refused to participate in the NES. (0=no; 1=yes) Political Information Discuss Politics Number of Calls Persuasion Letter Refusal Conversion Dummy variable indicating whether a persuasion letter was sent to the respondent. (0=no; 1=yes) Dummy variable indicating the race of the interviewer. (0=white interviewer; 1=Black interviewer) Interviewer Black 45 APPENDIX B: CODING PROTOCOL, 1989 NYC ELECTION VARIABLES Age Black Age of the respondent, in years. Dummy indicating the race of the respondent. Excludes black Hispanics. (0=non-black; 1=black) Dummy indicating whether the respondent is Hispanic. Excludes black Hispanics. (0=non-Hispanic; 1=Hispanic) Dummy indicating whether the respondent is a black Hispanic. (0=non-black Hispanic; 1=black Hispanic) Dummy indicating the gender of the respondent (0=male; 1=female) 6 category NES education variable measuring highest level of education (0=grade school; 1=advanced degree). A series of 7 dummy variables indicating the respondents reported income, or if the respondent’s income was not ascertained. The omitted category is an income of $30,000-$49,999 A series of 5 dummy variables indicating the religion in which the respondent was brought up (Protestant, Catholic, Jewish, other religion, and refused to identify religion). The omitted category is respondents who are Protestants. Dummy indicating if the respondent is registered with the Republican Party (0=not registered as a Republican; 1=registered as a Republican) Dummy indicating if the respondent is registered with the Democratic Party. This category of party identification is the omitted category in the analyses. (0=not registered as a Democrat; 1=registered as a Democrat) Dummy indicating if the respondent is registered as an independent (0=not registered as an independent; 1=registered as an independent) Hispanic (White and Other) Black Hispanic Female Education Income Variables Religion Variables Republican Democrat Independent 46 APPENDIX B: CODING PROTOCOL, 1989 NYC VARIABLES (CONTINUED) Other Party Identifier Dummy indicating if the respondent is registered with a party other than the Democratic Party or the Republican Party (0=not registered with another party; 1=registered with another party) Dummy indicating self-identification as “very liberal” or “liberal” on a five-point ideology scale. (0=not liberal; 1=liberal) Dummy indicating self-identification as “very conservative” or “conservative” on a five-point ideology scale. (0=not conservative; 1=conservative) Dummy indicating self-identification as “moderate” on a five-point ideology scale. (0=not moderate; 1=moderate) Dummy indicating respondents who “don’t know” their ideology or “don’t think in those terms” and decline to place themselves on a five-point ideology scale. (0=claim ideology; 1=claim no ideology) Variable indicating how likely the respondent thinks it is that he or she will vote in the mayoral election. If the respondent thinks that they will “probably vote,” “chances are 50/50,” or they “don’t think they will vote” this variable is scored a “0.” If the respondent is “certain to vote,” the variable is scored a “1.” Dummy variable indicating whether the respondent voted in the 1988 Presidential election. (0=No; 1=yes) Liberal Conservative Moderate No Ideology Certainty of Voting Voted in 1988 47 REFERENCES Achen, Christopher H. 1975. “Mass Political Attitudes and the Survey Response.” American Political Science Review. 69:1218-1231. ————. 1986. The Statistical Analysis of Quasi-Experiments. Berkeley: University of California Press. Anderson, Barbara A., Brian D. Silver, and Paul R. Abramson. 1988. “The Effects of Race of the Interviewer on Race-related Attitudes of Black Respondents in SRC/CPS National Election Studies.” Public Opinion Quarterly 52:289-324. Bartels, Larry M. 1988. Presidential Primaries and the Dynamics of Public Choice. Princeton, NJ: Princeton University Press. ————. 1996. “Uninformed Votes: Information Effects in Presidential Elections.” American Journal of Political Science, 40:194-230. Berinsky, Adam J. 1997. “The Miracle Revisited: The Micro-Foundations of Aggregate Public Opinion.” Paper presented at the 1997 Annual Conference of the American Political Science Association, Washington, DC. ————. 1998. “Towards an Integrated Theory of the Survey Response: The Mediated Communication Model.” Department of Political Science, University of Michigan. Typescript.. Blumer, Herbert. 1948. “Public Opinion and Public Opinion Polling.” American Sociological Review 13:542-554. Breen, Richard. 1996. Regression Models: Censored, Sample Selected, or Truncated Data. Sage University Paper series on Quantitative Applications in the Social Sciences, 07-111. Thousand Oaks, CA: Sage. Brehm, John. 1993. The Phantom Respondents: Opinion Surveys and Political Representation. Ann Arbor: University of Michigan Press. Carmines, Edward G. and James A. Stimson. 1980. The Two Faces of Issue Voting. American Political Science Review 79: 78-91. Chong, Dennis. 1993. “How People Think, Reason, and Feel about Rights and Liberties,” American Journal of Political Science, 37:867-99. ————. 1996. “Creating Common Frames of Reference on Political Issues.” In Diana C. Mutz, Paul M. Sniderman, and Richard A. Brody (eds.) Political Persuasion and Attitude Change. Ann Arbor: University of Michigan Press. Clymer, Adam. 1989. “Election Day Shows What the Opinion Polls Can’t Do.” New York Times, November 12. 48 Converse, Philip E. 1964. “The Nature of Belief Systems in The Mass Publics,” in David Apter (ed.) Ideology and Discontent. New York: Free Press. ————. 1970. “Attitudes and Non-Attitudes: Continuation of a Dialog,” in Edward R. Tufte (ed.) The Quantitative Analysis of Social Problems. Reading, MA: Addison-Wesley Publishing Company. Dovidio, John F. and Russell H. Fazio. 1992. “New Technologies for the Direct and Indirect Assessment of Attitudes.” in Judith M. Tanur (ed.) Questions about Questions: Inquiries into the Cognitive Bases of Surveys . New York: Russell Sage Foundation. Dubin, Jeffrey A. and Douglas Rivers. 1989. Selection Bias in Linear Regression, Logit and Probit Models. Sociological Methods and Research 18, no. 2&3: 360-90. Feldman, Stanley and John Zaller. 1992. "The Political Culture of Ambivalence." American Journal of Political Science 36: 268-307. Feldman, Stanley. 1988. "Structure and Consistency in Public Opinion: The Role of Core Beliefs and Values." American Journal of Political Science 32: 416-440. Finkel, Steven E., Thomas M. Guterbock, and Marian J. Borg. 1991. “Race-of-Interviewer Effects in a Preelection Poll: Virginia 1989.” Public Opinion Quarterly 55:313-330. Ginsberg, Benjamin. 1986. The Captive Public. New York: Basic Books. Greene, William H. 1997. Econometric Analysis, Third Edition. New York: MacMillian Publishing Company. Greene, William H. 1995. LIMDEP Version 7.0 User’s Manual. New York: Econometric Software, Inc. Heckman, James J. 1979. “Sample Selection Bias as a Specification Error.” Econometrica. 47:153-161. Herbst, Susan. 1993. Numbered Voices: How Opinion Polling Has Shaped American Politics. Chicago: University of Chicago Press. Hochschild, Jennifer. 1984. The New American Dilemma: Liberal Democracy and School Desegregation. New Haven: Yale University Press. Hout, Michael, Clem Brooks and Jeff Manza. 1995. "The Democratic Class Struggle in the United States: 1948-1992." American Sociological Review 60:805-828. Huckfeldt, Robert and John Sprague. 1987. Citizens, Politics, and Social Communication: Information and Influence in a Presidential Campaign. New York: Cambridge University Press. Hurley, Norman. 1997. “Do People really Feel What They Tell Us in Surveys: Toward Unobtrusive Measures of Racial Attitudes.” Paper presented at the 1997 Annual Conference of the American Political Science Association, Washington, DC. Iyengar, Shanto and Donald Kinder. 1987. News that Matters. Chicago: University of Chicago Press. 49 Jackman, Mary R. 1978. “General and Applied Tolerance: Does Education Increase Commitment to Racial Integration?” American Journal of Political Science 22:302-324. Jackman, Simon. 1997. “Compulsory Voting in Australia.” Paper presented at the 1997 Annual Conference of the American Political Science Association, Washington, DC. Keith, Bruce E., David B. Magleby, Candice J. Nelson, Elizabeth Orr, Mark C. Westlye, and Raymond E. Wolfinger. 1992. The Myth of the Independent Voter. Berkeley: University of California Press. Kinder, Donald R. 1983. "Diversity and Complexity in American Public Opinion." in Ada Finifter, ed., Political Science: The State of the Discipline. ————. 1998. “Opinion and Action in the Realm of Politics.” in Daniel Gilbert, Susan Fiske, and Gardner Lindsey (eds.) Handbook of Social Psychology, fourth edition. Boston: McGraw Hill. Kinder, Donald R. and Lynn M. Sanders. 1990. “Mimicking Political Debate with Survey Questions: The Case of White Opinion on Affirmative Action for Blacks.” Social Cognition 8:73-103. ————. 1996. Divided by Color: Racial Politics and Democratic Ideals. Chicago: Chicago University Press. Krosnick, Jon A. and Michael Milburn. 1990. “Psychological Determinants of Political Opinionation.” Social Cognition 8:49-72. Kuran, Timur. 1995. Private Truths, Public Lies: The Social Consequences of Preference Falsification. Cambridge: Harvard University Press. Madalla, G. S. 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press. McConnell, Scott. 1990. “The Making of the Mayor 1989.” Commentary. 48: 29-38. Noelle-Neumann, Elisabeth. 1984. The Spiral of Silence: Public Opinion — Our Social Skin. Chicago: University of Chicago Press. Reeves, Keith. 1997. Voting Hopes or Fears? White Voters, Black Candidates and Racial Politics in America. New York: Oxford University Press. Rosenthal, Andrew. 1989. “The 1989 Elections: Predicting the Outcome; Broad Disparities in Votes and Polls Raising Questions.” New York Times, November 9. Schuman, Howard and Jean M. Converse. 1971. “The Effect of Black and White Interviewers on Black Responses.” Public Opinion Quarterly 35:44-68. Schuman, Howard, Charlotte Steeh, Lawrence Bobo, and Maria Krysan. 1997. Racial Attitudes in America: Trends and Interpretations, Revised Edition. Cambridge, MA: Harvard University Press. Schuman, Howard and Stanley Presser. 1981. Questions and Answers in Attitude Surveys. New York: 50 Academic Press. Singer, Eleanor, Dawn R. Von Thurn, and Esther R. Miller. 1995. “Confidentiality Assurances and Response: A Quantitative Review of the Experimental Literature.” Public Opinion Quarterly 59:66-77. Snyder, Mark. 1983. “The Influence of Individuals on Situations: Implications for Understanding the Links between Personality and Social Behavior.” Journal of Personality. 51:497-516. Stimson, James A. 1991. Public Opinion in America: Moods, Cycles, and Swings. Boulder: Westview Press. Stimson, James A., Michael B. MacKuen, and Robert S. Erikson. 1995. “Dynamic Representation.” American Political Science Review. 89:543-565. Sudman, Seymour Norman M. Bradburn and Norbert Schwarz. 1996. Thinking About Answers: The Application of Cognitive Processes to Survey Methodology. San Francisco: Jossey-Bass Publishers. Tourangeau, Roger. 1987. “Attitude Measurement: A Cognitive Perspective.” In Hans J. Hippler Norbert Schwarz and Seymour Sudman (ed.) New York: Springer-Verlag. ————. 1992. Context Effects on Responses to Attitude Questions: Attitudes As Memory Structures. in Norbert Schwarz and Seymour Sudman (eds). Context Effects in Social and Psychological Research. New York: Springer-Verlag. Tourangeau, Roger and Kenneth Rasinski. 1988. “Cognitive Processes Underlying Context Effects in Attitude Measurement.” Psychological Bulletin. 103:299-314. Traugott, Michael W. and Vincent Price. “Exit Polls in the 1989 Virginia Gubernatorial Race: Where Did They Go Wrong?” Public Opinion Quarterly 56:245-253. Verba, Sidney. 1996. “The Citizen as Respondent: Sample Surveys and American Democracy.” American Political Science Review 90:1-7. Verba, Sidney, Kay Lehman Schlozman and Henry E. Brady. 1995. Voice and Equality: Civic Voluntarism in American Politics. Cambridge, MA: Harvard University Press. Wilson, Timothy D. and Sara D. Hodges. 1992. Attitudes As Temporary Constructions. in. Leonard L. Matin and Abraham Tesser (ed.). The Construction of Social Judgments Hilsdale, NJ: Lawrence Erlbaum Associates. Zaller, John. 1992. The Nature and Origins of Mass Opinion. New York: Cambridge University Press. Zaller, John and Stanley Feldman. 1992. "A Simple Theory of the Survey Response." American Journal of Political Science 36: 579-616. 51