Party Affiliation, Partisanship, and Political Beliefs: A Field Experiment PDF

American Political Science Review Vol. 104, No. 4 November 2010...

American Political Science Review Vol. 104, No. 4 November 2010 doi:10.1017/S0003055410000407 Downloaded from https://www.cambridge.org/core. University of Maryland College Park, on 06 Mar 2018 at 15:50:54, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055410000407 Party Affiliation, Partisanship, and Political Beliefs: A Field Experiment ALAN S. GERBER, GREGORY A. HUBER, and EBONYA WASHINGTON Yale University P artisanship is strongly correlated with attitudes and behavior, but it is unclear from this pattern whether partisan identity has a causal effect on political behavior and attitudes. We report the results of a field experiment that investigates the causal effect of party identification. Prior to the February 2008 Connecticut presidential primary, researchers sent a mailing to a random sample of unaffiliated registered voters who, in a pretreatment survey, leaned toward a political party. The mailing informed the subjects that only voters registered with a party were able to participate in the upcoming presidential primary. Subjects were surveyed again in June 2008. Comparing posttreatment survey responses to subjects’ baseline survey responses, we find that those reminded of the need to register with a party were more likely to identify with a party and showed stronger partisanship. Further, we find that the treatment group also demonstrated greater concordance than the control group between their pretreatment latent partisanship and their posttreatment reported voting behavior and intentions and evaluations of partisan figures. Thus, our treatment, which appears to have caused a strengthening of partisan identity, also appears to have caused a shift in subjects’ candidate preferences and evaluations of salient political figures. This finding is consistent with the claim that partisanship is an active force changing how citizens behave in and perceive the political world. S cholars from a variety of disciplines contend that and Huber 2010; Lupia 1992; Rahn 1993; Zaller 1992), allegiances and group affiliations, from national- preferences for biased political information (Lau and ism and religious identities to ethnic and kinship Redlawsk 2001; Redlawsk 2002), and the persistence ties, powerfully affect attitudes and behavior. One such over time of partisan affiliations (Alwin and Krosnick identity is partisanship, which political scientists have 1991; Green, Palmquist, and Schickler 2002; Jennings hypothesized is an active force shaping how individ- and Niemi 1974; Niemi and Jennings 1991). Across ac- uals evaluate and interact with the political world (in counts, both political and beyond, a common thread is the United States: Campbell et al. 1960; more recently, the claim that affiliations and identities cause the out- Bartels 2002; abroad: Brader and Tucker 2001; Dancy- comes associated with holding a particular allegiance. gier and Saunders 2006; Whitefield and Evans 1999). The claim that party identification is more than a Evidence presented to support the importance of par- summary of political attitudes or a “standing decision” tisanship includes the strong correlation between par- regarding candidate choice, but instead might play a tisanship and political opinions (vote choice: Bartels causal role in attitude formation, is consistent with the 2000; Campbell et al. 1960; Fiorina 1981; Miller large body of work in social psychology demonstrating 1991; assessments of the economy: Bartels 2002; the power of social identification to alter attitudes and Erikson 2004; Wlezien, Franklin, and Twiggs 1997), behavior. According to social identity theorists, it is the divergence among conflicting partisans in inter- a common human tendency, perhaps evolutionary in pretations of common events (Bartels 2002; Gerber origin, for individuals to distinguish between in-groups, those to which they belong, and out-groups (Sumner 1906). Belonging leads to formation of a group-based Alan S. Gerber is Professor, Department of Political Science, social identity that includes emotional attachments to Institution for Social and Policy Studies, Yale University, 77 the group and a tendency to favor the in-group (Tajfel Prospect Street, P.O. Box 208209, New Haven, CT 06520-8209 1978; Tajfel and Turner 1986). Individuals who perceive ([email protected]). Gregory A. Huber is Associate Professor, Department of Political themselves as members of a group may also internalize Science, Institution for Social and Policy Studies, Yale University, the group’s norms and values and use these as a guide 77 Prospect Street, P.O. Box 208209, New Haven, CT 06520-8209 for their own attitudes and behaviors (Brewer and ([email protected]). Brown 1998). Following Weisberg and Greene (2003), Ebonya Washington is Associate Professor, Department of Eco- nomics, Yale University. P.O. Box 208264, New Haven, CT 06520- among others, and applying this logic of social identity 8264 ([email protected]). formation to partisanship suggests that identifying with This research was paid for by a grant from the Yale University a party may be akin to forming a social identity as a Institution for Social and Policy Studies. member of that party and, as a consequence, may cause We thank Conor Dowling, David Doherty, and Bryan Gervais the individual to adopt the party’s values and develop for research assistance. An earlier version of this article was pre- sented at the 2009 Annual Meeting of the Midwest Political Science more favorable attitudes toward the party’s candidates Association, as well as at seminars at Columbia University, Cornell and causes.1 University, the University of North Carolina, Georgetown Univer- sity, Harvard University, the University of Michigan, the University of Pennsylvania, and Yale University. We thank participants in those 1 Weisberg and Greene suggest that “simply categorizing oneself forums, as well as the coeditors and reviewers, for helpful feedback politically as sympathetic to either party is likely enough to begin and advice. Any remaining errors are, of course, our own. psychological group processes” (Weisberg and Greene 2003, 90). 720  American Political Science Review Vol. 104, No. 4 Downloaded from https://www.cambridge.org/core. University of Maryland College Park, on 06 Mar 2018 at 15:50:54, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055410000407 There is substantial empirical support for the general variable: Bartels 2000; Bartels 2002; Carsey and Lay- theoretical claim that group influences and group mem- man 2006; Goren 2005; using state registration laws as bership can cause attitudes and behavior. Numerous an exogenous factor: Burden and Greene 2000), we studies suggest that the pressure to conform to group identify in existing research several persistent threats norms, especially when the behavior in question may be to unbiased measurement of causal effects. Thus, al- observed by others, can have large effects on individ- though previous research has shown that measures of ual behavior (e.g., Asch 1951; Cialdini 2001; Cialdini partisanship have large and robust predictive power and Goldstein 2004; Gerber, Green, and Larimer in statistical models of a variety of political outcomes, 2008). Previous research supports the specific claim we argue that existing studies have not demonstrated that an individual’s social identity leads to in-group bias that these relationships reflect the causal influence of (Mullen, Brown, and Smith 1992). For example, exper- these affiliations. In the next section, we review prior imental work on social identity shows that increasing contributions to the literature on the causal effect of the salience of in-group versus out-group distinctions partisanship and describe how our approach differs can increase in-group bias (see Yamagishi et al. 2008) from previous efforts. and, in the context of politics, discrimination along eth- We perform an experiment to assess whether parti- nic lines appears to increase when elections make those sanship has a causal effect on political attitudes. Al- ethnic ties more salient (Michelitch 2010). In this vein, though we employ a field experiment and focus on a both national identities and self-categorization (e.g., different form of group identity, our design parallels when checking particular boxes that describe one’s the basic approach used in the laboratory by psycholo- racial and ethnic identity on the census) may have a gists studying the effects of group affiliation using the possible role in shaping subsequent opinions. minimal group approach. We employ a randomized One of the most compelling demonstrations of the intervention that generates a group affiliation (in this power of social identification is found in research test- case, partisanship) and we then examine the effect of ing the “minimal-group paradigm” (Tajfel et al. 1971). this affiliation on attitudes and reported behaviors (in In this line of research, laboratory experiments are this case, political attitudes and behaviors). Our experi- performed that begin by creating groups based on the ment was fielded in the state of Connecticut during the most trivial and contrived differences, such as merely 2008 presidential primary election season. Although informing subjects that they have been randomly as- all registered voters may vote in any general election, signed to a group. After establishing group identities, Connecticut has a closed primary system in which only researchers then observe subsequent attitudes and be- voters affiliated with a party can vote in that party’s havior. They find that, among other things, subjects give primary. We performed a pretreatment survey of a in-group members more positive ratings than equiva- sample of registered independents (those who were lent out-group members (Brewer 1979; Brown, Tajfel, not formally registered with either the Democratic or and Turner 1980; Mullen, Brown, and Smith 1992) and Republican party and who also indicated they did not resource allocations are biased toward members of the already consider themselves Democrats or Republi- in-group (Tajfel et al. 1971). cans) and, based on their response to an item that Returning to the specific focus on partisanship, al- asked which party the respondent felt closer to, we though many previous studies undertake to measure classified some respondents as “latent” Democrats or whether and to what extent partisan identities shape Republicans (alternatively, in the language of Keith et political views, there are some important limitations al. 1992, these are “closet partisans”). Treated individ- to existing empirical approaches. A persistent concern uals received a mailing reminding them of the need to regarding existing research is the possibility that the ob- register with a party to be permitted to participate in served correlation between partisanship and politically that party’s upcoming presidential preference primary. relevant outcomes may originate in unobserved factors Four months after our intervention, we returned to the that are correlated with (or cause) both partisan iden- field to survey subjects about their partisanship and tities and beliefs (Bartels 2000; Fiorina 2002). Further, other attitudes. These data were supplemented with causality may flow in both directions, with partisan- information about party registration and turnout gath- ship reflecting political attitudes and events as well as ered from the Connecticut voter file. causing them (Allsop and Weisberg 1988; Beasley and We found that treated latent partisans, who by def- Joslyn 2001; Brody and Rothenberg 1988; Converse inition identified with neither party just months prior, 1976; Fiorina 1981; Franklin and Jackson 1983; Kessel were more than seven percentage points more likely to 1968; MacKuen, Erikson, and Stimson 1989; Norrander identify with their previously latent parties than those and Wilcox 1993; Weisberg and Smith 1991). Although assigned to the control group. Treatment group re- scholars have implemented a variety of research ap- sponses to the standard seven-point party identification proaches in an attempt to disentangle correlation from scale also became more polarized. Our intervention is causation (using lagged partisanship as an independent the first that we are aware of to induce partisan feelings over long periods of time and outside of the laboratory setting. (The intervention also increased party registra- For a discussion of social identity theory and partisanship, see the tion with the party of one’s latent partisanship by more valuable survey by Weisberg and Greene (2003), which includes a discussion of prior applications of social identity theory to party than eight percentage points and turnout in the primary identification, including Abrams (1994), Greene (1999), Hogg and election by more than four percentage points.) We then Abrams (1988), and Kelly (1988, 1989). employ this randomly induced partisanship to test key 721  Partisanship and Political Beliefs November 2010 Downloaded from https://www.cambridge.org/core. University of Maryland College Park, on 06 Mar 2018 at 15:50:54, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055410000407 theoretical arguments about the role of partisanship in Unobserved heterogeneity will bias estimates of β shaping political opinions and behaviors. We find that if there are any factors not included in M that are in addition to heightened partisan identities, treatment correlated with X and also affect Y. We label these group members were increasingly partisan in their vot- unmeasured factors U (for unmeasured variables). In ing choices and evaluations of partisan figures and in- most survey settings, factors in U include variables such stitutions. Thus, randomly induced variation in partisan as wealth, heredity, personality, educational and em- identities yields changes in attitudes and planned vot- ployment experience, and parental socialization. These ing decisions consistent with claims that partisanship sorts of variables are both hard to measure accurately is an active force shaping how citizens behave in the (even when attempts are made) and likely to have political world. Moreover, as there are many situations consequential effects on Y. Without including all the in a typical person’s daily life where the state or other factors that might plausibly affect Y and are also corre- organizations ask the individual to designate or affirm a lated with or cause partisanship (and therefore belong group identity, our findings suggest that it is worth con- in M but are instead left in U), analysis exploiting ob- sidering whether these procedures themselves might served variation in partisanship cannot rule out the affect people’s attitudes. alternative that partisanship (and therefore β) merely The remainder of this article is organized as follows. proxies for correlated but unmeasured factors. Con- First, we review the prior empirical literature and high- sistent with this concern, analysts regularly find that light the methodological concerns motivating our ex- including additional variables in M reduces the esti- perimental design, and then describe the experiment. mated effect of partisanship (β) on political opinions Next, we present results demonstrating that our treat- and behaviors (e.g., Fiorina 2002). ment altered partisan identity (as well as party regis- Endogenous partisanship also poses a threat to tration and voter turnout) and partisan views. Given causal interpretation of estimates of β. Regression the variety of behavioral and attitudinal effects arising analysis cannot distinguish the effect of X on Y from from our intervention, we also consider whether the the effect of Y on X. If Y is a measure of political prefer- causal impact of our experimental manipulation oper- ences, it is reasonable to anticipate that Y might affect ated through changes in party identification or through another choice, partisan identity. Thus, one cannot rule some other mediating mechanism. Some of our treat- out the possibility that it is instead opinions that cause ment effect estimates have large standard errors. In the partisanship. final section, we discuss some of the important limita- Returning to omitted variable bias, one approach de- tions of our analysis and conclude. signed to address concerns about unmeasured factors (U) that shape both partisanship (X) and the outcome of interest (Y) is employing panel data in which the RESEARCH DESIGN, CAUSAL INFERENCE, same respondent is interviewed multiple times (e.g., AND THE EFFECT OF PARTISANSHIP Bartels 2000). The relationship between changes in In this section, we discuss the barriers to causal infer- partisanship and changes in Y can then be used to ence in existing research and describe an alternative estimate β without bias originating in U, but this re- technique for measuring partisanship’s effects. The ear- quires the restrictive assumption that the change in liest and most common approach to demonstrating the the unmeasured factors is not related to the change effects of partisanship on political attitudes or behav- in partisanship.3,4 If U changes, or if the effect of U ior relies on cross-sectional data (or a pooled series of on X or Y varies, however, then β may still be biased. cross sections). Those data are then used to estimate a Thus, the plausibility of this research strategy hinges on regression of the following form: whether there are convincing explanations for why the respondent’s partisanship changed and why the sources Yi = α + βXi + γMi + ε, (1) of those changes would not also affect attitudes, a con- sideration typically neglected in this research. where Y is the outcome of interest for individual i (e.g., In practice, panel estimates of β are considered can- intended vote choice), X is partisanship, M is a vector of didates for causal interpretation when it is reasonable measured control variables (M for measured variables), to assume that the change in X (in this case partisan- and ε is the error term. There are many important stud- ship) is due to changes in some factor that does not ies that follow this estimation approach, including the directly affect Y. However, it is quite plausible that classic studies in The American Voter. However, analy- observed changes in partisanship over time are due sis employing this specification is vulnerable to biased to changes in factors that are unmeasured (U), such estimation of β, the effect of partisanship on outcome Y, in a variety of circumstances. The most important threats to inference originate in (a) omitted variable surement error may generate bias in estimates of β as unobserved heterogeneity if that measurement error is correlated with X. Cor- bias due to unobserved differences across individu- related measurement error in X and Y may also generate bias in als (unobserved heterogeneity) and (b) endogenous estimates of β. We detail particular cases in which measurement partisanship.2 error is a salient concern below. it+1 – Yit = α + β(Xit+1 – 3 Formally, this regression is of the form Y Xit ) + γMi + ε. A related but distinct approach discussed below is 2 Additional threats include measurement error in M. The effects the effect of lagged partisanship on change in attitudes. of measurement error are complex and depend on the covariances 4 Additionally, as in the cross-sectional approach, measurement er- among the variables and the pattern of measurement error. Mea- ror may also cause estimates of β to be biased. 722  American Political Science Review Vol. 104, No. 4 Downloaded from https://www.cambridge.org/core. University of Maryland College Park, on 06 Mar 2018 at 15:50:54, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055410000407 as life experiences (e.g., parenthood), wealth changes, but examine a different set of issue attitudes. Again, changes in the views of close friends and relatives, or the central idea is that if the change in values or atti- changes in religious beliefs, and any of these might tudes is predicted by partisanship in the first period, this cause changes in both partisanship (X) and Y. Alter- suggests that the change was caused by partisanship. natively, U may remain constant, but the nature of Although this strategy has appeal, it suffers from sev- political conflict might vary, thereby causing U to now eral potential limitations. First, a correlation between affect both X and Y differently. For example, in a polit- the change in the respondent’s measured attitudes and ical context where issues of taxation are more salient, lagged partisanship could arise due to measurement wealthier individuals might hold different policy views error. Suppose the political attitude (say, support for and feel closer to one party than when social issues are cutting taxes) is the product of the interaction between the focus of the political agenda.5 some stable underlying value toward government and The researcher may offer arguments as to why the measurement error due to survey design idiosyncrasies change in partisanship (or, in the cross-sectional ap- or the particular events in the news or on the political proach, the level of partisanship) is uncorrelated with agenda at the time of the survey (e.g., whether the the regression error. However, it is important to rec- issue of the day is a tax cut for the wealthy). Even if ognize that this lack of correlation is ultimately an the underlying value is stable, the measured attitude untested assumption. Moreover, even if this omitted at any point in time will be a noisy measure of the variable issue is overcome, the panel approach still underlying value and will also be correlated with the cannot resolve the uncertainty about the direction of lagged measured attitude, with each measured attitude causality. When the dependent variable is respondent affected by survey design and context effects. If parti- attitudes, it may be the case that changes in Y (attitudes sanship is also (partially) a function of the respondent’s toward political issues) cause variation in partisanship stable underlying value (i.e., if Republicans are on av- rather than the other way around. This difficulty is erage those who believe in smaller government), then observed in Jacoby (1988), which examines the rela- partisanship is another noisy measure of the under- tionship among the perceptions of party positions on lying value. Because regression analysis will typically issues, a respondent’s own issue attitudes, and strength minimize the prediction error in current measured at- of partisanship. Jacoby reports that the degree to which titudes by placing weight on both of the available noisy attitudes are predicted by the respondent’s perceptions measures of the underlying value (lagged measured of a party’s candidate’s positions increases with the attitudes and lagged partisanship), a regression of the respondent’s strength of party identification. This is current attitude on lagged partisanship and lagged at- consistent with the idea that stronger party identifi- titudes will place some weight on lagged partisanship cation produces stronger attitude agreement, as well even if partisanship does not have any effect on the as with the alternative explanation that respondents movement in the attitude over time.7 report being stronger partisans when they agree more Second, the regression estimates may be spurious with their party’s candidates. due to omitted variable bias. To consider the research Two important recent papers use survey data from design in a slightly simplified form, suppose that the multiple periods to assess the relationship between researcher performs a regression of the change in atti- party identification and political attitudes. In contrast tudes on lagged partisanship. The coefficient estimate to the panel approach, which examines the effect of on partisanship from this bivariate regression will cap- change in partisanship on change in attitudes, this work ture the effect of lagged partisanship on the change measures whether change in attitudes between two in attitudes, as well as the net effect of all omitted points in time is predicted by the respondent’s par- variables that are correlated with partisanship and that tisanship measured at the first point in time.6 Because also predict the change in values. For example, parti- partisanship is measured prior to the attitude change, sans tend to have discussion networks with similar par- this approach avoids the danger that an observed cor- tisans (Huckfeldt and Sprague 1987). If more Demo- relation between changes in partisanship and changes cratic (Republican) social environments tend to move in attitudes is driven by attitude changes generating a respondents in a more liberal (conservative) direction, change in partisanship. Goren (2005) uses this design this will produce an upward bias in the estimated ef- and finds that political values (such as an index mea- fect of partisanship if the model does not account for suring support for limited government) are predicted these discussion patterns. Similarly, if respondents’ life by lagged partisanship when controlling for lagged experiences during the period are correlated with their measures of the value, a finding that is interpreted as partisanship (which is plausible given that partisans evidence that partisanship causes a change in values. Carsey and Layman (2006) employ a similar approach 7 Formally, suppose that the respondent’s true underlying value is Z. Fix Z and partisanship (X) as stable. The respondent’s measured attitudes, Yt+1 and Yt , are a mix of Z and period-specific error terms 5 Note also that in the absence of an explanation for observed produced by, among other things, survey artifacts and variations in changes in partisanship (X), there is little reason to believe that political context at times t and t + 1. Measured partisanship Xt is also the changes in X cause variation in Y. Rather, changes in X may a function of Z, equal to Z plus a random error. A prediction of Yt+1 reflect common shocks to X and Y originating in U or measurement will place weight on both Yt and Xt , but this does not show that X is error in X, Y, or M. causing Yt+1 (or Z). In the case of independent errors, the weight on it+1 – Yit = α + βXit + γMi + ε or Yit+1 = α + 6 Formally, this is Y one of the past measures is increasing in the error in measuring the δYit + βXit + γMi + ε. other. 723