Explaining Politics, Not Polls Reexamining Macropartisanship with Recalibrated Nes Data

Like all surveys, the American National Election Studies (NES) imperfectly reflects population characteristics. There are wellknown differences between actual and NES-reported turnout rates and between actual and NES-reported presidential vote divisions. This research seeks to determine whether the aggregate misrepresentation of turnout and vote choice affects the aggregate measurement of party identification: macropartisanship. After NES data are reweighted to correct for turnout and vote choice errors, macropartisanship is found to be more stable, to be less sensitive to short-term political conditions, and to have shifted more in the Republican direction in the early 1980s. The strength of partisanship also declined a bit more in the 1970s and rebounded a bit less in recent years than the uncorrected NES data indicate. According to the American National Election Studies (NES), 77.6 percent of eligible Americans turned out to vote in the 2008 presidential election. Barack Obama received 54.8 percent to John McCain’s 45.2 percent of the two-party popular vote. This is what the NES data indicate, but this is not what happened. By our best counts, turnout was 61.7 percent of the voting-eligible population, Obama received 53.7 percent, and McCain received 46.3 percent of the two-party popular vote. Despite inaccuracies of this sort, the NES remains the “gold standard” of survey data to scholars of American elections. NES data are routinely reported JAMES E. CAMPBELL is Professor and Chair of the Department of Political Science, University at Buffalo, State University of New York, Buffalo, NY, USA. The data used in this study were made available by the American National Election Studies (NES) with funding since 1977 from the National Science Foundation. I wish to thank the NES, the NSF, and the Survey Research Center and the Center for Political Studies at the University of Michigan for collecting and distributing these data. I also thank Congressional Quarterly and Michael McDonald and the United States Election Project for collecting and making available the vote division and turnout data that were used in this study. *Address correspondence to James E. Campbell, Department of Political Science, 520 Park Hall, University at Buffalo, SUNY, Buffalo, NY 14260, USA; e-mail: [email protected]. Public Opinion Quarterly, Vol. 74, No. 4, Winter 2010, pp. 616–642 doi: 10.1093/poq/nfq042 Advance Access published on September 23, 2010 © The Author 2010. Published by Oxford University Press on behalf of the American Association for Public Opinion Research. All rights reserved. For permissions, please e-mail: [email protected] at SU N Y B ufalo on Sptem er 9, 2012 http://poqrdjournals.org/ D ow nladed from as what the electorate thinks and does, what affects that thinking and behavior, and how the electorate’s thinking and behavior has changed over time (Abramowitz 2004; Abramson, Aldrich, and Rohde 2005; Jamieson 2000, 14, 82; Keith et al. 1992, 14; Lewis-Beck et al. 2008; Martinez and Gill 2005). However, like all survey data, and as the 2008 example illustrates, NES data are imperfect. Sampling error, nonresponse rates, measurement errors, framing effects, question-wording sensitivities, respondent misreporting, and various other problems produce errors in surveys that make them less than perfectly representative of political reality (Clausen 1968–69; Crespi 1988; Asher 2007). In general, students of public opinion and voting behavior assume that these errors are small and do not significantly affect their analyses. The general question posed in this article is whether the known inaccuracies of NES studies regarding aggregate turnout and presidential vote choice, the two most important electoral behaviors, significantly affect the NES accounts of other aggregate-level political attributes related to turnout and vote choice. In particular, do the known errors in aggregate turnout and vote choice appreciably affect the NES’s aggregate measure of party identification, macropartisanship? Since there are well-established associations between the strength of party identification and the decision to turn out to vote, and between party identification and the vote choice at the individual level (Campbell et al. 1960; Lewis-Beck et al. 2008), and parallel associations between aggregate partisanship and both the national vote division (Converse 1966; Abramson and Ostrom 1991; Erikson, MacKuen, and Stimson 2002) and turnout levels (Abramson and Aldrich 1982; Rosenstone and Hansen 2003), there are strong reasons to suspect that errors in the distribution of turnout and the presidential vote would also generate errors in the aggregate distribution of partisanship. More specifically, we should expect that whichever party is overrepresented in the presidential vote is also overrepresented in party identification, and that the general underrepresentation of nonvoters in NES data underrepresents weaker partisans and independents who are less likely to vote. Put differently, since most partisans vote for their party’s presidential candidate, if voters for a candidate are underrepresented among survey respondents, then the partisans who voted for that candidate are probably also underrepresented. These remain suspicions, however, because unlike aggregate turnout and the presidential vote division, there is no external benchmark for aggregate partisanship that makes the errors in its measurement obvious. This study tests these suspicions by reweighting NES data to bring them into line with the actual aggregate benchmarks of 1. Clausen (1968–69, 591) also observes that a portion of the discrepancy between NES measures and the actual electorate is a function of NES’s definition of the “sample universe” as something other than all eligible voters. The NES’s sample universe excludes a number of potential voters (e.g., hospital patients, nursing home residents, hotel residents). 2. This study concerns only the accuracy of survey measurements of turnout, vote choice, and partisanship, and not the causal relationships among these variables. Macropartisaship 617 at SU N Y B ufalo on Sptem er 9, 2012 http://poqrdjournals.org/ D ow nladed from the presidential vote and turnout distributions, in order to measure more accurately macropartisanship and its change over time. The analysis makes four points. First, though often regarded as small in absolute terms and within the normal bounds of sampling error, NES’s errors with respect to presidential vote choice and turnout are not trivial, because the scale of real change in American politics also involves only “small-scale shifts around the middle” (Stimson 2004, 159). Second, aggregate presidential vote choice errors are highly variable from one election to the next. This distorts over-time estimates of aggregate political change. Third, both the turnout and presidential vote choice data errors can be corrected (or at least mitigated) by reweighting the data to the known population parameters of turnout and the presidential vote choice. Such corrections are routinely performed to correct for the underor overrepresentation of demographic characteristics (race, gender, etc.) and of political variables in exit polls (McDonald 2007). Fourth, the corrected NES data, brought into line with the known population parameters of turnout and presidential vote choice, produce significant changes in the distribution and dynamics of macropartisanship. These changes produced by the corrections in NES data have important implications for two debates about macropartisanship. The first concerns the stability of macropartisanship and its sensitivity to short-term political conditions. Both Erikson, MacKuen, and Stimson (1989; 2002) and BoxSteffensmeier and Smith (1996) find macropartisanship to be quite changeable and sensitive to the recent performance of the parties. On the other hand, Abramson and Ostrom (1991) and Green, Palmquist, and Schickler (2002) find macropartisanship to be stable and largely insensitive to shortterm conditions. The second debate concerns the existence of long-term or realigning changes in macropartisanship. Erikson, MacKuen, and Stimson (2002) find that macropartisanship changes continually, as in a “moving equilibrium,” rather than through the punctuated change of realignments. This view challenges the conventional wisdom about realignments rooted in the early work of Key (1955; 1959) and Burnham (1970) and the later research by Meffert, Norpoth, and Ruhil (2001) that found a realignment in macropartisanship in the 1980s. The following sections take the analysis through three steps. First, the extent of the NES errors in turnout and vote choice is examined. Second, the assumptions required to adjust or reweight NES data are identified, and the weights are calculated and applied to the data. Third, the distribution and dynamics of party identification in the corrected and uncorrected NES data are compared. Details about the NES data and the questions used in this study may be found in Appendix 1. Campbell 618 at SU N Y B ufalo on Sptem er 9, 2012 http://poqrdjournals.org/ D ow nladed from