Changing Parties or Changing Attitudes?: Uncovering the Partisan Change Process

Transcription

1 Changing Parties or Changing Attitudes?: Uncovering the Partisan Change Process Thomas M. Carsey* Department of Political Science University of Illinois-Chicago 1007 W. Harrison St. Chicago, IL and Geoffrey C. Layman Department of Political Science Vanderbilt University Box 8262, Station B Nashville, TN geoffrey.c.layman@vanderbilt.edu * Authors listed alphabetically. Prepared for delivery at the 1999 Annual Meeting of the Midwest Political Science Association, April 15-17, 1999, Chicago.

2 Changing Parties or Changing Attitudes?: Uncovering the Partisan Change Process A central question in political science concerns the impact of powerful new political issues that cut across the lines of established partisan cleavages (see Key 1955, 1959; Campbell 1966; Burnham 1970; Sundquist 1983; and Carmines and Stimson 1989 for some of the major theoretical treatments of the partisan change process). Of particular interest are the mechanisms that lead to new partisan cleavages, or the polarization of the parties electoral coalitions along the lines of the new issues. Some scholars focus on party conversion, arguing that change results primarily from large numbers of individuals discarding old party attachments for new ones (Erikson and Tedin 1981; Sundquist 1983; Burnham 1970). The existing partisan alignment is based upon a dominant set of issues; individuals choose their party attachment based on their views on those issues. However, when a powerful new set of issues comes along and cuts the electorate in ways different from the old issues, many members of a party may find themselves in greater agreement with the stands of the other party on those issues than with those of their own party, and thus may change their party loyalties accordingly. Others argue that aggregate partisan change stems primarily from mobilization, with large numbers of previously inactive or unattached voters being aroused by the new issues and choosing a party affiliation based on their positions on those issues (Andersen 1976, 1979; Clubb, Flanigan, and Zingale 1980; Salisbury and MacKuen 1981). A third group devotes attention to generational replacement, in which older voters who came of political age before the emergence of the new issues are replaced by younger voters for whom the new issues are more salient and who thus choose their party affiliations based on them (Beck 1979; Carmines and Stimson 1989; Key 1959). What these proposed mechanisms of partisan change all have in common is that positions on the new issues are assumed to drive party affiliations. Either existing partisans switch parties in order to bring their party affiliations in line with their stances on the new issues, or new voters choose their party affiliations for the first time based on the new issues. None of these ideas suggest that partisanship drives issue positions, or that it is change in individual issue positions rather than party identification that leads to mass party polarization. The same is generally true of spatial theories of partisan change (Aldrich 1983a, 1983b; Carmines 1994), along with most versions of the general spatial model (e.g. Enelow and Hinich 1984; Hinich and Munger 1997). These theories typically assume that issue preferences among citizens are constant (at least in the short run), and that candidates or parties pick locations within the given issue space of the electorate hoping to win majority support. Even theories that posit change in the issue space of the electorate generally focus on altering the salience of various policy dimensions (Schattschneider 1960; Riker 1990; Carsey 1999) or in providing additional information (Alvarez 1997) rather than actually converting voters on an issue. Like the realignment literature noted above, what is assumed to change is the willingness of voters to support a particular candidate or political party and not the issue preferences that voters hold (though Carsey 1999 notes some exceptions). In this paper, we contend that these accounts of the partisan change process are incomplete because they fail to consider the possibility that change in the aggregate issue positions of the parties mass coalitions may result also from existing partisans changing their positions on the issues. In other words, as the positions of a party s leaders, candidates, and platforms on a set of issues moves to a more ideologically-extreme position, individual party members may convert on the issues, bringing their own attitudes into line with those of their party. That many individuals might change their position on an issue because the position of their party has changed may seem unlikely when one considers the nature of the issues associated with long-term partisan change. Unlike most political issues, toward which most citizens are apathetic and on which attitudes are weak and unstable (cf. Converse 1964; Converse and Markus 1979), realigning issues or issue evolutions arouse the passions of the mass public. They tend to be easily understood, to produce strong emotions, and to be associated with more deeply-held attitudes (Sundquist 1983; Carmines and 1

3 Stimson 1989). Indeed, attitudes on the two sets of issues most closely associated with recent partisan change racial issues in the 1960s and early 1970s (Carmines and Stimson 1989) and cultural issues like abortion and women s rights in the 1980s and 1990s (Carmines and Layman 1997; Adams 1997) have been shown to be noticeably more stable than attitudes on other issues (Converse 1964; Converse and Markus 1979). However, three bodies of literature point to attitudinal conversion in response to party affiliation as a strong possibility. The first is the extensive research on the role of party identification in individual voting decisions. The American Voter (Campbell et al. 1960) argued that party identification is a deeplyheld psychological attachment to a political party. It is formed largely during pre-adult socialization and is unlikely to change much thereafter. Thus, party identification acts as a conceptual filter, leading individuals to view the issues and candidates in a certain way, and ultimately exerting a substantial impact on the vote. Party identification is viewed as largely exogenous to issue attitudes and voting behavior. It has a considerable influence on issue-attitudes and candidate evaluations, but these short-term factors have little effect on it. Subsequent research argues that party identification is not wholly exogenous. Instead, it does change in response to issue-attitudes, candidate evaluations, evaluations of party performance, and even voting decisions (Jackson 1975; Page and Jones 1979; Markus and Converse 1979; Fiorina 1981; Franklin and Jackson 1983; Franklin 1984; Brody and Rothenberg 1988). However, none of this revisionist research argues that party identification does not continue to exert a substantial influence on evaluations of issues and candidates. Even if it is somewhat endogenous, party identification s effect on individual s issue positions and candidate evaluations are at least as great, if not greater, than the impact of those factors on it (Markus and Converse 1979). Analysis of panel data has found individual party identification to be more stable over time than evaluations of political figures or attitudes toward any political issues (Converse and Markus 1979). When random measurement error is corrected, party identification is found to be even more stable and almost entirely exogenous to issue, candidate, and performance evaluations in the short-run (Green and Palmquist 1990, 1994). If party identification has the kind of effect on individual-level issue-attitudes that this body of research suggests, then, even though attitudes on emotional issues associated with long-term partisan change are deeply held, many individuals who are already aligned with a party may still change their attitudes on such issues to bring them closer to the changing stance of their party. Also supporting the idea of attitude change in response to partisan loyalties is the work by John Zaller (1992) on elite-driven change in public opinion. Zaller contends that predispositions such as partisanship and liberal-conservative ideology are more fundamental and less likely to change in the short-run than are attitudes on particular issues. Thus, when political elites from opposite ends of the ideological spectrum or from the Democratic and Republican parties take polarized stands on an issue, as they do in the case of a realigning issue, the result is a polarization effect in public opinion. Rather than changing their core predispositions, such as their party affiliations, in response to elite change on the issue, individuals, particularly the most politically-attentive ones, move their own positions on the issue closer to the polarized position of the elites who share their political predispositions. A third area of research that supports the notion of attitudinal conversion is that on partisan change among party activists. There is considerable evidence that activists convert on issues as the positions of party leaders and platforms change, and that this conversion contributes significantly to aggregate partisan change (Miller and Jennings 1986; Stone, Rapoport, and Abramowitz 1990; Rapoport and Stone 1994; Herrera 1995). Moreover, recent research shows that even on an issue as emotional as abortion, there has been substantial conversion among activists, and that conversion has made a major contribution to overall party polarization on the issue (Layman and Carsey 1998; Carsey and Layman 1999). Party activists tend to have more deeply-held attitudes on issues than do ordinary voters (Herrera 1992) and increasing numbers of activists are motivated primarily by their positions on issues (Aldrich 1995). Thus, if activists convert on issues as the positions of their parties change, there is every reason to expect that ordinary party identifiers may do so as well. 2

4 The analysis in this paper assesses the extent to which aggregate change in the positions of the parties mass coalitions on a new set of issues results from attitudinal conversion due to party identification as opposed to partisan conversion due to attitudes on the new issues. We examine this from the perspective of the issues that have been associated most recently with substantial changes in party politics: cultural issues such as abortion, women s rights, homosexual rights, and prayer in the public schools. Since their emergence on the political scene in the late 1960s and early 1970s, these issues have become a focus of growing partisan conflict. Party polarization on these issues has grown substantially at nearly all levels of politics: in the roll-call votes of Democratic and Republican members of Congress (Adams 1997), in the attitudes of delegates to the parties national conventions (Layman 1999a), in the party platforms adopted by those conventions (Layman 1999b), in the attitudes of the parties campaign activists (Layman and Carsey 1998), and in the aggregate positions of the parties mass electoral coalitions (Carmines and Layman 1997). Moreover, this growth in partisan cultural differences is not something that occurred years ago and is now subsiding. Instead, the decade of the 1990s has witnessed increases in party polarization on cultural issues that are as large or larger than those that occurred in the 1980s (Adams 1997; Carmines and Layman 1997; Layman 1999a, 1999b). 1 Clearly these issues have created and continue to create significant changes in partisan politics. They are also issues on which attitudes are relatively stable over time (Converse and Markus 1979). Thus, they provide a particularly rigorous test of the idea that attitudinal conversion on new issues contributes to partisan change. The most direct way to sort out whether individuals are changing their party identifications in response to their cultural issue-attitudes or are changing their cultural attitudes in response to their party affiliations is to use panel data. So, we begin by examining the reciprocal relationship between party identification and issue-attitudes over time with the two panel studies conducted by the National Election Studies (NES) in the 1990s: the panel and the panel. 2 There are, however, two potential problems with this panel analysis. First, the panels only cover a six-year period in the history of issues that have been on the political state for two or three decades albeit the six-year period over which the largest increases in mass party polarization have occurred. Second, these panels may be too short to pick up any significant amount of individual-level change in either party identification or cultural issue-attitudes, especially given the relative stability of both. Traditionally, there has been no solution to this second problem other than longer panel studies. However, the method of Ecological Inference (EI) recently developed by King (1997) provides the potential for gauging individual-level 1 Taking the abortion issue as an example, the difference in the mean scores of Republican and Democratic identifiers in the National Election Studies on a four-point abortion scale, ranging from most pro-choice to most pro-life, increased from -.05 in 1980 to.08 in 1990 and to.36 in The difference in the mean scores of Republican and Democratic national convention delegates on a similar four-point scale (taken from Layman (1999a) for and from the CBS News/The New York Times polls of national convention delegates for 1996) increased from.55 in 1980 to.94 in 1988 and 1.32 in The difference in the percentage of congressional Democrats and Republicans supporting the pro-choice position in roll-call votes on abortion (taken from Adams (1997) for 1980 through 1994 and computed by the authors for 1995 and 1996) increased from 30.8 in 1980 to 47.1 in 1989 and to 60.2 in The NES also conducted panel studies in 1972, 1974, and 1976 and over the course of the 1980 campaign. These panels are not conducive to examining partisan change on cultural issues. Since the parties and their leaders had not yet taken clear stands on issues like abortion in the early 1970s, there was little change in the aggregate positions of the Democratic and Republican coalitions over this period. The surveys in 1972, 1974, and 1976 also contained no questions about school prayer or homosexual rights and the 1974 survey did not contain a question about abortion. Abortion was the only cultural issue asked about in more than one wave of the 1980 panel and the question wording on abortion changed between the second and third waves, making it impossible to gauge individual-level change. 3

5 change from cross-sectional surveys. 3 We apply this method to the presidential-year and midterm-year NES cross-sectional studies from 1972 to 1996 to gain a sense of the extent of individual-level change among Democratic and Republican identifiers on the abortion issue, the issue the forefront of contemporary cultural politics and the cultural issue appearing most consistently in the NES surveys. The combination of panel analyses over a short period and ecological analysis of cross-sectional data covering a much longer period should provide a thorough look at the question of the extent to which party-driven attitudinal conversion drives aggregate partisan change. Attitudinal Conversion vs. Partisan Conversion in the NES Panels The increase in party polarization on cultural attitudes 4 in the 1990s that is evident in the NES cross-sections is equally as evident in the NES panel studies. 5 The question is did this increase in partisan differences occur because people changed their party affiliations based on their cultural attitudes, because people changed their cultural attitudes based on their party identifications, or both? Table 1 takes a first look at this question by showing the mean cultural attitudes in two different waves of the panel (1990 and 1992 or 1992 and 1996) for individuals who kept the same three-category party identification between the two waves and the mean party identifications in two different waves of the panel (1990 and 1992 or 1992 and 1996) for individuals who remained in the same category of a trichotomized measure of cultural attitudes between the two waves. Even at the individual level, there is clear evidence of party polarization over these two panels. Between 1990 and 1992, individuals who were Democrats in both years became, on average, more liberal in their cultural attitudes while individuals who were Republicans in both years grew more conservative. The same phenomenon is evident between 1992 and Individuals who maintained a Democratic party identification converted, on average, to a more liberal cultural attitude while individuals who were affiliated with the GOP in both years converted to a more conservative position on cultural issues. Given the overall range of the cultural issue scales, none of these changes are particularly large. However, the fact that there is some noticeable change in mean cultural attitudes over very short periods and that the changes are in opposite directions for Democrats and Republicans provides support for the notion that one of the factors creating partisan polarization on powerful new issues is individual conversion on the issues among partisans. Table 1 also uncovers evidence of individuals changing their party identifications in response to their attitudes on cultural issues. Individuals who had liberal cultural views as defined by the trichotomized scale in both 1990 and 1992 became, on average, slightly more Democratic in their party affiliation while individuals who had conservative cultural attitudes in both years became slightly more 3 Penubarti and Schuessler (most recently 1998) have produced a series of papers applying the same approach to the study of change in presidential approval. We follow their basic approach here. 4 Our measure of cultural attitudes in the panel is the sum of respondents scores on questions questions about the circumstances under which abortion should be legal, parental consent for abortion, and government funding of abortions (the school prayer and women s role questions were only asked to separate halves of the sample in 1990 and there were no questions asked about homosexual rights). They range from zero (most liberal) to nine (most conservative) and have a reliability coefficient (alpha) of.60. Cultural attitudes in the panel are the sum of respondents' scores on questions about abortion, laws to protect homosexuals against discrimination, homosexuals in the military, the role of women in society, and school prayer. They range from zero (most liberal) to 18 (most conservative) and have a reliability coefficient of In the panel, the mean cultural attitudes in 1990 were 4.69 for Democrats and 5.43 for Republicans while the mean cultural attitudes in 1992 were 4.19 for Democrats and 6.22 for Republicans. In the panel, the mean cultural attitudes in 1992 were 5.32 for Democrats and 7.08 for Republicans while the mean cultural attitudes in 1996 were 5.11 for Democrats and 7.51 for Republicans. 4

6 supportive of the GOP. The patterns for cultural liberals and cultural conservatives were the same between 1992 and 1996, but the growth of Democratic ties among liberals and of Republican ties among conservatives is larger over this four-year period. In sum, this table suggests that the polarization of the parties coalitions on cultural issues resulted both from party identifiers bringing their cultural attitudes closer in line with the polarized stances of their parties leaders and candidates and from individuals with polarized cultural views moving their partisan loyalties closer to the party that best represents those views. It is, of course, true that the estimates of change for both party identification and cultural issueattitudes may be contaminated by random measurement error. For various reasons inexact question wording or response options, interviewer mistakes, coding errors, or top-of-the-head answers by respondents a survey response can never be counted on to be an error free measure of an underlying attitude or trait. Failure to correct for measurement error in panel data can lead to the appearance of changes in attitudes or orientations when no real change has occurred. In fact, recent research shows specifically that when measurement error is corrected, party identification (Green and Palmquist 1990, 1994) and issue-attitudes (Krosnick 1991) are considerably more stable than they appear to be when no corrections are made. Since it is random, measurement error should not lead to a systematic bias in the direction of the change we have uncovered: Democrats and Republicans moving in opposite directions over time on cultural issues, cultural liberals and cultural conservatives moving in opposite directions over time on the party identification scale. But, it may well have led us to overestimate the amount of polarizing change. In order to take measurement error into account and examine the relationship between party identification and cultural attitudes, we estimate a covariance structure model that includes both a measurement model and a structural model that proposes cross-lagged effects between party identification and cultural attitudes over time. The model is presented in Figure 1. Identifying the model requires at least three waves of panel data, so we use the NES panel. There is only one observed measure of party identification in the NES surveys: the standard seven-point party identification scale. So, the measurement model for party identification is the standard single-indicator measurement model in which observed party identification (in rectangles) in each wave of the panel is a function of latent or true party identification (in ovals) and some random measurement error (E). The standard set of assumptions for the measurement errors and the structural disturbance terms (D) is that the covariances between the measurement errors and the latent variables and between the measurement errors and the disturbance terms are all zero (Bollen 1989; Finkel 1995). However, even with these constraints, the model remains underidentified because there are six independent pieces of information the variances and covariances of the observed indicators of party identification in the three waves of the panel with which to estimate 11 parameters: the variances of the three measurement errors, the variances of the three disturbance terms, the effects of latent party identification on observed party identification in each of the three waves, and the two stability coefficients (the effects of latent party identification in 1992 and 1994 on latent party identification in 1994 and 1996, respectively). In order to identify the measurement model for party identification, we employ a standard set of restrictions proposed by Wiley and Wiley (1970). We assume that the error variances of observed party identification are equal over time and that the effects of observed party identification on latent party identification are all equal to one. The latter constraint also ensures that unobserved party identification has the same scale as the observed seven-point indicator of party identification. In contrast to party identification, there are three observed indicators of cultural attitudes that are present in each of the three waves of the panel: a four-point scale of abortion attitudes, a seven-point scale of attitudes toward women s role in society, and a four-point scale of attitudes toward prayer in the public schools. Having multiple indicators allows us to place fewer restrictions on the cultural attitudes measurement model. We estimate separate variances for the measurement errors for each of the nine observed indicators. We also allow the measurement errors for the indicators of attitudes on the same issue to covary over time. We set the factor loading for observed abortion attitude in each of the three 5

7 panel waves to one so that the cultural attitudes factor has the same four-point scale as the abortion indicator. However, the factor loadings for the women s role and school prayer indicators are estimated in each wave rather than fixed. We also assume that the stability coefficients the effect of unobserved cultural attitudes at one time point on cultural attitudes at the next time point are equal between 1992 and 1994 and 1994 and The structural portion of the model proposes that there are cross-lagged effects between true party identification and true cultural issue-attitudes over time: An individual s cultural issue-attitude at one time point is a function of his or her cultural attitude at the previous time point and his or her party identification at the previous time point. An individual s party identification at one time point is a function of his or her party identification and cultural issue-attitude at the previous time point. 6 Because the model controls for both variables values at the previous time point, the lagged effects of party identification on cultural attitudes and of cultural attitudes on party identification can be interpreted as the effect of one variable on changes in the other variable over time (Finkel 1995). 7 We model the relationship between cultural issue attitudes and partisanship as reciprocal, but not simultaneous, for three reasons. First, and most importantly, our substantive focus differs from that of the voting behavior literature where non-recursive models of party identification and issue-attitudes are generally found (e.g Page and Jones 1979). We are not interested in whether party identification is exogenous or endogenous to issue-attitudes at a single point in time. Instead, we are interested in the effect of party identification on changes over time in cultural attitudes and of cultural attitudes on changes over time in party identification. The cross-lagged-effects model is better equipped to examine that than is a model of contemporaneous effects. Second, Finkel (1995) demonstrates that the cross-lagged model s applicability is not limited to discrete time processes of change: for example, party identification in 1992 affects change in cultural attitudes between 1992 and Instead, it is appropriate even if we assume that the reciprocal effects between variables occur continuously over time. Under these circumstances, the cross-lagged model tends not to be misleading about the direction of causal influence (Dwyer 1983, 352). Finally, the estimation of simultaneous-effects models requires the use of instrumental variables that are exogenous to the variables in the model and that have direct causal effects on only one of the endogenous variables. These assumptions can often be quite dubious, particularly with survey data (Dwyer 1983; Finkel 1995). Moreover, if there is measurement error in the instruments one uses, one s estimates may be seriously biased (Green and Palmquist 1990). Table 2 presents the estimates, obtained with AMOS 3.6, of the factor loadings, stability coefficients, and cross-lagged effects in our model. 8 The estimated variances and covariances of the measurement errors are presented in Appendix A. As noted earlier, the effects of latent party identification on observed party identification are constrained to equal one in each panel wave. The 6 Following standard practice in the estimation of cross-lagged-effects models (cf. Finkel 1995), we assume that the cross-lagged effects between cultural attitudes and party identification between 1992 and 1994 are equal to the cross-lagged effects between the two variables between 1994 and The cross-lagged structure also meets the requirements of a Granger causality test, which states that a variable Granger-causes another if any value of the first variable at a prior time point has a significant effect on the second variable at the current time point, controlling for the second variable s prior values. 8 We also estimated models that included controls for a number of sociodemographic variables race, gender, income, education, region of residence, and religious affiliation as well as models that also included attitudes toward social welfare issues and racial issues, both corrected for measurement error through a multipleindicator measurement model, and proposed cross-lagged effects between those attitudes and party identification. The inclusion of these variables did not cause noticeable changes in the estimated reciprocal effects between party identification and cultural attitudes, nor did they significantly improve the overall fit of the model. 6

8 factor loadings for observed abortion attitude in each wave are also set to one, while we estimated the factor loadings for observed attitudes toward women s role in society and prayer in the public schools. Given that abortion is the most-publicized and perhaps the most emotional of the contemporary cultural issues, it is not surprising that the abortion indicator loads most strongly on the cultural attitudes factor in both 1994 and 1996, with attitude toward women s role in society loading most strongly in The fact that the school prayer indicator has the weakest loading in each year is also not surprising since school prayer is not the subject of as intense a conflict as that surrounding the other cultural issues. Overwhelming majorities of Americans over 85 percent in the 1992, 1994, and 1996 NES surveys favor some form of prayer in the public schools, with strong pluralities or narrow majorities favoring the opportunity for silent prayer over all other options. Like past research, we find that when we correct for measurement error, party identification is highly stable over time. Its standardized stability coefficients are both over.9, indicating that an individual s party identification in one election year is a very good predictor of his or her party identification in the next election year. Also similar to past research is our finding that cultural attitudes are quite stable from one election year to the next. Their standardized stability coefficients are not as large as those for party identification (.70 and.81), but they still indicate an impressive amount of stability. Moving next to the reciprocal relationship between partisanship and cultural issue attitudes, we find that cultural issue-attitudes do change over time in response to the effects of prior party identification. Individuals with stronger ties to the Republican party are more likely than more- Democratic citizens to convert to more conservative cultural attitudes over time. Meanwhile, cultural attitudes at one time point also lead to changes in party identification between that time point and the next. Cultural conservatives are more likely than cultural liberals to move their party affiliation in a Republican direction over time. Although statistically-significant, neither of the cross-lagged effects is large. A one-unit increase in identification with the Republican party leads to an increase in cultural conservatism of only.03 on a four-point scale. On average, strong Republicans (seven on the party identification scale) only convert.18 scale points more in a culturally-conservative direction over a two-year period than do strong Democrats (one on the party identification scale). A one-unit increase in cultural conservatism leads to an increase in Republican identification of only.15 on the seven-point scale. On average, individuals with the most conservative cultural attitudes (four) convert only.45 scale points more toward the Republican party over a two-year period than do individuals with the most culturally-liberal attitudes (one). However, two things should be kept in mind. First, both cultural attitudes and party identification are deeply-held, highly-stable orientations. We would not expect them to change a great deal over a twoyear period, particularly when random measurement error has been removed. Second, these estimates are only for a two-year period. Over a longer period, they represent more substantial changes in both party identification and cultural attitudes. Moreover, we are not suggesting that a critical-election realignment on cultural issues occurred during the period from 1992 to Rather, we, like other scholars (cf. Adams 1997), propose that there has been a gradual polarization much like Carmines and Stimson s (1989) issue evolution model of partisan change along the lines of cultural issues. Our findings clearly demonstrate that this polarization occurred not just because people changed their partisan ties in response to their positions on cultural issues, but also because people changed their attitudes on cultural issues in response to their party affiliations. Assessing the Stability of Abortion Attitudes With Ecological Inference Methods In this section we extend our analysis to estimates of attitudinal change across a longer time period. Following Penubarti and Schuessler (1998), we apply King s (1997) solution to the ecological inference problem can to cross-sectional surveys taken at two points in time to generate estimates of individual-level change. Here we focus specifically on the degree of individual-level change that has taken place on abortion between 1972 and

9 Figure 2 plots the proportion of respondents to the various cross-sections of the NES that we classify as being pro-life and pro-choice on abortion. 9 Each score is based on responses to a single question with four possible answers. We classify those answering with one of the two more restrictive options as pro-life and those opting for the two less restrictive options as pro-choice. Note that the form of the question changed beginning in the 1980 survey, which likely accounts for the noticeable drop in the proportion of respondents classified as pro-life. 10 Aside from the dip resulting from this change, figure 2 demonstrates a remarkable aggregate stability in the public s attitudes toward abortion. From 1972 to 1978, between 56 and 58 percent of respondents chose the two most restrictive options on the abortion question. From 1980 to 1996, that figure ranged from 39 to 47 percent. This finding of aggregate stability is not new and has been replicated with other data sets (e.g. Adams 1997). One is tempted after looking at figure 2 to conclude that substantial individual-level stability must underlie this observed aggregate stability. We know that voters hold more consistent attitudes on abortion than just about any other issue (Converse and Markus 1979). In fact, the simple correlation on the four-point abortion scale for panel respondents in the NES panel is.62 (.59 for the dichotomous measure). Similarly, for the , and the panels, the same correlations are.77,.67 and.71 respectively (.68,.68, and.63 for the dichotomous measures). While fairly high, these correlations are not so high as to preclude change on abortion from taking place. In the panel study, fully 20 percent of respondents switched from one side of the abortion issue to the other. The comparable figures from the , and panels are 17 percent, 16 percent, and 18 percent, respectively. The aggregate stability in abortion attitudes shown in figure 1 suggests that switching on abortion occurs in both directions at roughly equal rates. However, this does not mean that such conversion happens in both directions at the same rate within each party. Thus, the increased polarization on abortion among party identifiers may be influenced by attitudinal change on abortion. Our analysis of the NES panel study indicates support for this conclusion. Here we wish to further explore the degree to which individual-level change has taken place on abortion across a longer time period as well as whether such change is associated with partisanship. Because panel data over longer periods of time does not exist, we turn to King s (1997) method of ecological inference to estimate rates of individual change on abortion over time. King s (1997) method is designed to estimate the parameters such as those presented in Table 3. Following Penubarti and Schuessler s (1998) notation, ecological data generally provides information on A 1 i, defined here as the proportion of respondents who are pro-life at time point 1 (e.g. 1994), and A 2 i, the proportion of respondents who are pro-life at time point 2 (e.g. 1996), (and thus 1-A 1 i and 1-A 2 i, the proportion of pro-choice respondents at t1 and t2). Ecological data does not supply direct measures of the individual-level parameters β 11 i, the proportion of pro-life respondents at t1 who remain pro-life at t2, or β 01 i, the proportion of pro-choice respondents at t1 who are pro-life at t2. In the absence of true panel data, King s method begins by breaking aggregate data down to subaggregate units. For King, that means reducing district-level voting statistics to the precinct level. For us was in The only cross-sectional NES survey from 1972 to 1996 in which the abortion question was not asked 10 From 1972 to 1978, the four response options on abortion were abortion should never be permitted; abortion should be permitted only if the life and health of the woman is in danger; abortion should be permitted if, due to personal reasons, the woman would have difficulty in caring for the child; and abortion should never be forbidden, since one should not require a woman to have a child she doesn t want. From 1980 to 1996, the four response options were by law, abortion should never be permitted; the law should permit abortion only in cases of rape, incest, or when the woman s life is in danger; the law should permit abortion for reasons other than rape, incest, or danger to the woman s life, but only after the need for the abortion has been clearly established; and by law, a woman should always be able to obtain an abortion as a matter of personal choice. 8

10 (and Penubarti and Schuessler), it requires that we dis-aggregate each cross-sectional survey into a set of sub-aggregates, denoted by the subscript i. We do so by constructing a series of demographic profiles. In this analysis, our profiles are constructed based on four variables: gender (male and female), education (high school graduate or less, more than high school graduate), income (lower third, middle third, upper third), and religious tradition (evangelical, mainline or black Protestant, Catholic, secular or other). 11 This produces 48 profiles, or sub-aggregate units of analysis ( = 48). To present an example, one profile would consist of male, high school graduate or less, middle income, Catholics. For each of these profiles, the values for A 1 and A 2 can be observed. 12 A 1 i and A 2 i measure the upper and lower bounds of β 11 i and β 01 i for each profile, and must by definition be between 0 and 1. In most empirical examples, however, they will fall within much narrower ranges. King s method incorporates these bounds into a statistical model to estimate the probable location of β 11 i and β 01 i within their known deterministic bounds. For every profile, the relationships defined in Table 3 imply the following equation: β i 01 = (A i 2 /(1-A i 1 )) - (A i 1 /(1-A i 1 ))β i 11 Because the values for A 1 i and A 2 i are known for every profile, a value of β 01 i could be computed for every hypothetical value of β 11 i. As a result, all possible values of β 11 i and β 01 i for each profile can be represented as a line. A figure that plots each possible line for each profile King refers to as a tomography plot. Being able to reduce the possible values that β 11 i and β 01 i could take on to a set of lines narrows the possible estimates of the values of β 11 i and β 01 i for each profile. King then generates statistical estimates of β 11 i and β 01 i assuming that the parameters follow a truncated bivariate normal distribution. Once the estimates of β 11 i and β 01 i for each profile are obtained, they can be used to estimate the aggregate parameters β 11 and β 01 for the entire sample. 13 To demonstrate the usefulness of this method, we first compare the estimates derived from EI to those which can be derived from the various NES panels. 14 To test the accuracy of the EI method, we treat each wave of the panel as if it were merely a cross-sectional survey. We generate 48 profiles based 1 on our four demographic variables for each wave. For each profile, we then compute the values for A i and A 2 i, which are the proportion of respondents in each profile that are pro-life in 1972 and 1976, respectively. This allows us to estimate β 11 i and β 01 i for each profile, which ultimately allows us to estimate the proportion of respondents that did and did not change their views on abortion from 1972 to 1976 without taking advantage of the panel nature of the data. 11 Penubarti and Schuessler (1998) report that their findings are robust to the selection of variables used to construct these profiles. We have found that to be true as well regarding our analysis of abortion, as we considered several other profiles using basic demographic variables, all of which produced the same basic results. 12 Penubarti and Schuessler (1998) correctly note that these can only be observed up to the level of sampling error. This suggests constructing profiles in a manner that does not produce many profiles that have a small number of respondents. Our experience suggests that profiles containing 10 or more respondents are desirable, which we achieve here. 13 Interested readers should consult King (1997) for a complete treatment of the method of ecological inference employed here along with Cho (1998) for a critique of King s method. 14 In some of the EI analyses that follow, we uncovered evidence of aggregation bias affecting the parameter estimates. King argues that EI estimates are robust to aggregation bias, but it is also easily corrected. See King (1997, pp 183-4) for a discussion of the approach. All table entries presented here are corrected for aggregation bias when it was observed. 9

11 Figure 3 presents a tomography plot for this estimation. Each straight line represents the possible combinations of β 11 i and β 01 i for a specific profile. Figure 2 also presents two confidence level curves which highlight the areas of the tomography square in which the point estimates for β 11 i and β 01 i fall. These intervals suggest that fairly large values of β i are expected, with somewhat lower values of β i predicted. The results presented in Figure 3 are summarized in the aggregate in Table 4. The entries in bold are the actual levels of individual-level change observed in the panel data. 15 Table 4 shows that of those who held a pro-life view on abortion in 1972, 83 percent continued to hold that view in Of those who were pro-choice on abortion in 1972, almost 25 percent of them switched to the pro-life view. Table 4 reports that the EI estimates of these two proportions (β 11 and β 01 ) are quite accurate. The relatively small standard errors illustrate the degree of certainty regarding these parameters. For each parameter, the true value is well within plus or minus two standard errors of the EI estimates. Figure 4 presents the posterior distribution of the estimates for β 11 and β 01 with the true value of these parameters indicated. Figure 4 illustrates graphically what Table 2 shows numerically the EI estimates are close to the known true values and the standard errors associated with those estimates are relatively small, suggesting fairly precise estimates of change. We also conducted similar analyses comparing abortion attitudes reported by respondents in the , , and waves of the NES panel study. In order to save space, we present only the tabular estimates. Table 5 paints a similar picture regarding the accuracy of the method. In nearly every case, the estimated parameters are within plus or minus two standard errors of the true value. In summary, our application of King s EI method to the analysis of change in abortion attitudes confirms Penubarti and Schuessler s analysis of change in presidential approval the method appears to accurately replicate the known individual level of change making use of only aggregate cross-sectional information. Having established the apparent accuracy of the method, we move now to an analysis of each cross-section of the NES from 1972 through 1996 in which the abortion question was asked. Because of the changes in question wording, we do not estimate change between 1978 and Table 6 presents the results. For each year, the method estimates that typically between 20 and 25 percent of individuals changed their view on abortion from one survey period to the next, which in all but one case constitutes a two-year interval. This is a substantial amount of attitudinal instability which might not be expected given the overall level of aggregate stability portrayed in Figure 2. Comparing the results presented in Tables 4 and 5 based on panel data with the results presented in Table 6 shows that the estimates in Table 6 remain quite accurate compared to the known levels of switching that took place in the panel. While within or near two standard deviations for the and periods, the EI estimates presented in Table 6 appear to over-estimate the level of change compared to what was observed at the individual level in the panel. We suspect this results from the sampling error inherent in survey research or more specifically from measurement error. However, even if sampling and/or measurement error contribute to slightly higher estimates of change than would be observed, there remains strong evidence that substantial position switching on abortion is taking place Here we do not account for measurement error as King s EI method is not well suited to do so. Thus, here and throughout this section, we likely over-estimate the rate of change on abortion. 16 A counter-argument to this point would be that panel participants, by having been asked the question previously, may be somewhat less inclined to report a changed position on the abortion issue in later waves of the study. If so, then the observed levels of change in the panel studies, taken here to be estimates of the true levels of change, may in fact underestimate the degree of change that is taking place in the large population. In other words, the EI estimates of change may be more accurate than those derived from individual-level panel studies. 10

12 We conclude this section with an examination of whether or not the estimated rates of conversion on abortion differ among partisans. We approach this question in two ways. First, we pooled the profiles from 1980 through 1996 together in a single data set containing 384 profiles. We next generated EI estimates for this entire pool. 17 We also calculated the proportion of respondents for each profile that were self-identified Republicans at T1 and T2. Following King s suggestion (p. 170), we regressed our profile-level estimates of β 11 i and β 01 i, respectfully, on the difference in the proportion of respondents that were Republican identifiers at T2 compared to T1 for each profile. The results, presented in Table 7, indicate that for groups that became more Republican in the aggregate, both β 11 i and β 01 i increased significantly. This finding means that, among individuals within profiles that became more Republican in the aggregate, those who were pro-life at T1 were more like to remain pro-life at T2 compared to those who were pro-life at T1 but were part of profile groups that became less Republican. Similarly, individuals who were pro-choice at T1 were more likely to convert to being pro-life at T2 as the profile group of which they were a part became more Republican. In our second approach to this question, we constructed new group profiles, this time using party identification instead of religious denomination as one of the profile characteristics. The new profiles were defined by gender (female or male), education (less than a high school graduate, high school graduate, more than high school graduate), income (bottom third, middle third, upper third), and party identification (Democrat, Independent, Republican). 18 This produces a possible 54 profiles. 19 The data for these profiles were then pooled from 1980 through 1996 in the same manner as previously described and a single analysis was conducted to produce another full set of EI estimates. 20 We next grouped the estimated based on the partisanship of the profiles and ran simple different of means tests to see whether the average estimates for β 11 i and β 01 i differed for Democrats, Independents, and Republicans. 21 Table 8 reports the results. It shows a statistically significant difference between all three partisan groupings for both β 11 i and β 01 i, though the substantive differences between Democrats and Independents is small. Table 8 shows again that Republicans who were pro-life at T1 are more likely to remain pro-life at T2 than are Democrats or Independents that were pro-life at T1. Similarly, Republicans that were pro-choice at T1 are more likely to switch to being pro-life at T2 than are either Democrats or Independents that were pro-choice at T1. To summarize, we find evidence of conversion on abortion in response to partisanship for the entire 1980 to 1996 time period using King s EI method. This evidence suggests that our findings 17 The aggregate estimates for β 11 and β 01 for the pooled data, with standard errors in parentheses, are.7478 (.0140) and.1858 (.0108), respectively. Independents. 18 Partisanship was measured using the standard 7-point scale, with independent leaners classified as 19 There were several instances in which one or two of the possible profiles had no observations for either T1 or T2, so the analysis reported here is often based on fewer than 54 profiles. This does not appear to influence the results. 20 The aggregate estimates of β 11 and β 01 for the pooled data for this set of profiles, with standard errors in parentheses, are.7230 (.0149) and.2051 (.0112). These estimates suggest slightly higher levels of conversion that those based on the original profile structure. We suspect that this reflects the replacement of religious denomination with party identification, which is more likely subject to change over time and possibly measurement error than is religious denomination, in the construction of the profiles. It may also be due to the slightly large number of profiles and/or a larger percentage of profiles having small numbers of respodents. 21 In conducting this analysis, each profile was weighted by the number of observations in it. 11