Macropartisanship A Replication and Critique

Macropartisanship: A Replication and Critique1 The discovery that citizens come to identify themselves with political parties and that these attachments persist even amid marked changes in the political environment Campbell, Converse, Miller, and Stokes 1960 ranks among the most important social science insights. Not only does the concept of party identi cation a ord the student of electoral politics a means by which to predict electoral behavior, it provides the basis for a causal explanation, one that traces voters' preferences and beliefs to enduring psychological attachments to parties as social groups. Political conduct and opinions, in e ect, have long term antecedents in the socialization experiences that in young adulthood crystalize into identi cations with groups and symbols. For decades, this conception of party identi cation has weathered a steady stream of empirical critiques attempting to show how partisanship changes in response to uctuations in the political fortunes of the parties. Evaluations of the parties' presidential candidates Page and Jones 1979, issues stances Franklin and Jackson 1983; Franklin 1992, and performance in o ce Fiorina 1981; Brody and Rothenberg 1988 have all been said to a ect party identi cation among adults. Although these ndings continue to enjoy wide currency within political science, they hinge on the unrealistic assumption that the Michigan party identi cation scale is an errorless measure of attitude. When allowance is made for even a small degree of measurement error, the empirical basis for these revisionist critiques is severely undermined Green and Palmquist 1990, 1994; Schickler and Green 1993-4, 1995.2 Party identi cation is by no means immune to change, but the circumstances under which one nds reliable evidence of change are by and large those anticipated by the authors of The American Voter. Party attachments can be disrupted in systems such as Canada in which new parties emerge that pit linguistic, ethnic, and regional identities against preexisting party identities Schickler and Green 1995 or in party systems such as contemporary Italy in which party labels are in ux. And the individuals most prone to change party identi cation or indeed a wide array of social identi cations; see Alwin and Krosnick 1991 1 We are grateful to Mike MacKuen, Jim Stimson, Bob Erikson, Renee Smith, and Janet Box-Ste ensmeier for graciously supplying us with data. We wish also to acknowledge ICPSR and the Roper Collection at Yale, both of which furnished survey data. This project was supported by grants from the Institution for Social and Policy Studies at Yale and the National Science Foundation, neither of which bear responsibility for any ndings we report. The data and analyses described in this paper are available from the authors on request. 2 The exception to this methodological generalization is Franklin and Jackson's 1983 analysis of issue proximity e ects, which instead relies on implausible identi cation restrictions see Schickler and Green 1993-4. 1 are those under the age of 30 Jennings and Niemi 1978; Green and Palmquist 1994. These ndings would suggest that in stable party systems generational replacement and secular change among young voters constitute the primary causes of change in the distribution of party attachments over time. The distribution of party a liations, like the distribution of religious or ethnic a liations, would be expected to move gradually as a result of population replacement. The partisan balance, while not altogether immutable, would be expected to adjust slowly to swings in political or economic fortunes and shift appreciably only when short-term forces favor one party for a sustained period of time. Party attachments, as Converse 1966 argues, function as ballast in an electoral system, stabilizing party competition amid shifting political currents. Since the mid-1980s, however, this hypothesis has been challenged by those who nd that the distribution of party identi cation not only shifts signi cantly over short periods of time, but moves in tandem with uctuations in presidential popularity, consumer sentiment, or objective economic indicators of unemployment and in ation MacKuen, Erikson, and Stimson 1989, 1992; Weisberg and Smith 1991. The evidence concerning macropartisan" change, or shifts in the proportion of Democrats and Republicans, is widely regarded as an important challenge to the view that partisan attachments are highly stable over time.3 This essay takes a closer look at the most important and in uential demonstration that macropartisanship adjusts to uctuations in the political environment, the work of MacKuen, Erikson, and Stimson 1989, 1992. Although these two Review essays have been widely cited and justly praised for their path-breaking use of aggregate data, their statistical analyses have not been examined closely.4 Reproducing and re ning As Green and Palmquist 1994 point out, ndings of aggregate volatility need not invalidate the argument that adults maintain the partisan identities over long periods of time and that the principal sources of macropartisan change are generational replacement and period e ects" among the young. Disaggregating the partisanship time-series by region suggests, for example, that most of the long-term slide in Democratic identi cation has occurred in the South, and much of that has been due to generational replacement Green and Schickler 1996. 4 The work that comes closest to replicating the MacKuen et al. analysis is the Cromwell et al. 1995 monograph, a methodology text that takes as its central example the causal interplay between consumer con dence and macropartisanship. But in reconstructing the MacKuen et al. analysis, the authors fail to take into account the fact that improving economic conditions have a positive e ect on the macropartisanship series during Democratic administrations and a negative e ect during Republican administrations. In e ect, Cromwell et al. evaluate a quite di erent proposition, namely, that hard times makes the public more Republican. 3 2 these results is important for several reasons. First, it illustrates the potential value of replication standards in social science cf. King 1995; Herrnson 1995; Maisel 1995. Here is an instance in which a signi cant empirical question hangs in the balance, where no complications arise due to the use of proprietary, con dential, or singular qualitative information, and where data gathering and analytic procedures are, in principle, entirely reproducible. Second, other time-series analyses of macropartisanship have produced con icting results. Ostrom and Abramson 1991, using the biennial National Election Study series, detect little movement in party identi cation or responsiveness to short-term forces. Whiteley's 1988 Box-Jenkins analysis of the 1984 NES Continuous Monitoring Survey failed to detect e ects of presidential and economic evaluations. Allsop and Weisberg 1988, on the other hand, nd a signi cant bivariate correlation between aggregate party identi cation and vote intention during this period, and Weisberg and Smith 1991 nd statistically signi cant links between partisanship and certain lags of unemployment, in ation, and presidential approval in monthly poll data from 1981-1986. Viewed against the backdrop of these other studies, the persuasiveness of the MacKuen et al. research derives from the fact that it traces short-term movements in partisanship using a much more extensive dataset, quarterly readings of party identi cation stretching over several decades. A nal reason to replicate the MacKuen et al. 1989 results concerns the sampling procedure used to generate their partisanship series. As MacKuen et al. explain p.1140, this series was derived from just a subset of all of the Gallup polls: The quarterly Gallup readings are themselves aggregated from bimonthly readings. Measures of macropartisanship were obtained from the Roper Center as a systematic sampling of party identi cation from the rst Gallup survey of every odd-numbered month. We reaggregated to quarters because key economic indicators are measured quarterly. After the publication of their essay, MacKuen et al. went on to gather data from all of the Gallup surveys. Through personal communication with the authors we learned that the original series used in MacKuen et al. 1989 no longer exists.5 The data can be reconstructed crudely using the authors' graph of the quarterly values p.1138. Because only 138 of 140 data points seem to have been plotted, ruler and compass cannot always discern the quarter associated with each macropartisanship reading. We can, however, reconstruct the variance of the entire series since this does not rely on the temporal sequencing of the data, estimated to be 20.92. This gure is almost exactly the value 20.93 implied by the summary statistics that MacKuen Using Roper Center documents, we attempted to recreate the series using the monthly Gallup polls and the sampling procedure used by MacKuen et al., but we were unable to reproduce the series graphed on p.1138. Because we cannot analyze the precise data used by MacKuen et al., our reanalysis is properly characterized as a replication rather than veri cation Herrnson 1985: 452. 5 3 et al. report on p.1136.6 The complete Gallup series, by contrast, has a variance of 16.98 over this same period. Apparently, 23 of the variance in the macropartisanship series used by MacKuen et al. 1989 is additional sampling variation. If the ndings reported in MacKuen et al. 1989 could not be reproduced using the complete series, the results obtained using the original series would become a statistical curiosity brought about by sampling. In the sections that follow, we use the updated Gallup series to replicate as closely as possible the original statistical analysis undertaken by MacKuen et al 1989. We then crossvalidate our results by replicating the analysis of CBS New York Times data undertaken by MacKuen et al. 1992. Last, we develop a more exible and parsimonious time series model linking approval, consumer sentiment, and macropartisanship. We use this revised model both to gauge the robustness of the MacKuen et al. results and to assess the degree to which consumer sentiment and presidential approval improve out-of-sample forecasts. None of our statistical analyses yields strong support for the proposition that macropartisanship uctuates appreciably in the wake of changes in economic performance or presidential popularity. The Measurement of Macropartisan Change Political scientists had become accustomed to thinking that the most noteworthy trend in the distribution of party identi cation was dealignment, or growing ranks of self-identi ed independents, when MacKuen, Erikson, and Stimson reshaped the terms of debate. Unlike previous students of American party a liation, these authors relied not on the placid series of biennial National Election Study surveys but instead on a quarterly compilation of surveys conducted by the Gallup organization.7 Besides charting party identi cation outside the con nes of the biennial election season, the authors sidestepped the dealignment issue. None of the more than 100,000 Gallup respondents who between 1953 and 1988 described themselves as independents makes it into MacKuen et al.'s macropartisanship time series. Their measure hinges solely on the relative numbers of Democrats and Republicans. Figure 1 traces the proportion of party identi ers who call themselves Democrats in 677 Gallup personal and 305 telephone surveys from 1953 to 1995, as well as the corresponding gures from the biennial National Election Surveys from 1952 through 1994. The solid line transforms the monthly Gallup surveys Taking the R2 that MacKuen et al. report in their Table 3 to be an adjusted R2 , we use the formula 2 2 e =1 , R  to obtain 20.93. The statistics that MacKuen et al. report in their Table N-1 suggest an even higher variance. We cannot account for this discrepancy. 7 Allsop and Weisberg 1988 argued that dynamics in party identi cation occurring within a campaign may be missed by the biennial National Election Study surveys but did not extend their empirical analysis to the broad period examined by MacKuen et al. 6 4 by aggregating them into quarters.8 The partisan balance, as gauged by the Gallup surveys, tips increasingly in the Democratic direction from the early 1950s until the mid-1960s. Some retrenchment occurs until the start of the Nixon Administration, when Democrats begin gradually to recapture their former dominance. With the election of Ronald Reagan, Democratic fortunes again wane. The Republican gains, though halted momentarily by the recession of 1982-3, continue through the rest of the Reagan Administration, so that the series ends at very nearly the same level at which it began.9 The quarterly series presented from 1953:1-1988:1 was provided by MacKuen, Erikson, and Stimson. Their quarterly data are based on personal interviews until 1988:1. Using data from individual in-person polls conducted during this period, we were able to reproduce this series with a high degree of accuracy. In 1985, Gallup had begun to phase out personal surveys, and in 1988 gaps begin to appear in the quarterly series of personal interviews. At this point, MacKuen et al. use a weighted average of personal and phone surveys, with a correction for the pro-Republican bias of Gallup telephone surveys. We use their series through 1988:1 when seeking to replicate their results. For extensions of their analyses, we substitute telephone surveys aggregated quarterly for the period 1988:2-1996:1. Abramson and Ostrom 1992 express concern about comparisons between Gallup telephone and in-person surveys, and in some ways our data attest to the important di erences between the two. Not only do Gallup telephone surveys contact more Republicans, but examination of the 36 quarters for which both personal and telephone measures of macropartisanship exist 1985:1-95:4 reveals just a .19 correlation between personal and telephone assessments of macropartisanship. Of the two, only the phone series correlates appreciably with the quarterly macropartisanship series constructed from CBS New York Times polls see below, so we have opted to update the 1953-88 Gallup series using phone polls. To correct for the pro-Republican bias of Gallup telephone polls, we add 4.525 to macropartisanship scores for the period 1988:2-1996:1, which is the average di erence between personal and phone interviews for the period 1985-1996 during quarters in which both types of Gallup surveys were conducted. The pattern of results we report is not sensitive to the particular manner in which the macropartisanship series is updated. 9 The volatility of the Gallup macropartisanship series in Figure 1 is partly cosmetic. Macropartisanship, as operationalized by MacKuen et al. 1989, is the proportion of all party identi ers who a liate with the Democrats. Respondents calling themselves independents are discarded from the computation of macropartisanship, which contributes to the apparent volatility of the series. Suppose, for example, respondents in Poll A are 34 Democrats, 33 Republicans, and 33 independents. In Poll B, 33 of all respondents call themselves Democrats, 34 Republicans, and 33 independents. This one percentage-point shift in the 8 5 MacKuen et al. regard these uctuations as strong prima facie evidence against the prevailing view that the partisan balance holds intact until suddenly and permanently altered by an electoral realignment. These movements in partisanship are often of a magnitude large enough to suggest electoral realignment. Note that these shifts are not temporary but persist from quarter to quarter. Yet they have nothing like the permanence envisioned in realignment notions. The partisan balance is not nearly so stable as The American Voter or critical realignment theory would lead us to expect. p.1128 As Abramson and Ostrom 1991 point out, however, swings in macropartisanship are much less evident in the National Election Study series. Since the mid-1960s, the biennial NES surveys have charted a gradual decline in Democratic identi cation. The NES trendline slices through the more volatile Gallup series; there is no sign, for example, of one of the most striking shifts in the Gallup series, the surge in Democratic a liation between 1968-1978. Although MacKuen et al. 1992:476 state reassuringly that virtually all NES and GSS General Social Survey observations are within the range of Gallup monthly observations for the same period," this can properly be said of just 13 of 20 NES surveys plotted in their diagram and 15 of 22 in ours.10 Why does there seem to be more movement in the Gallup series than the NES series? A lengthy list of possibilities suggest themselves. In one group are contrasts in sampling design and procedure; few commercial polls have ever approximated the care with which the National Election Study designs and implements its surveys. In another group are hypotheses about the e ects of interview context and question wording. The Gallup surveys occur more frequently and therefore may pick up month-to-month change in partisanship that would elude a biennial election survey. This feature of the Gallup series puts the partisan stability thesis to an especially demanding test, particularly as the Gallup polls often coincide with the emergence of a salient political issue or concern. One suspects that a respondent's partisan coloration might tend in the Democratic Republican direction after discussing at length the Iran-Contra Whitewater scandal during the course of a half-hour interview. Moreover, the Gallup polls have used a measure of party identi cation that is di erent from the traditional Michigan question. The Michigan item reads Generally speaking, do you usually think of yourself as a Republican, Democrat, Independent, or what?", whereas the Gallup measure asks In politics, as of today, do you consider yourself a Republican, a Democrat, or an Independent?"11 A great deal of controversy surrounds the issue of question wording. Abramson and Ostrom 1991 proportion of Democrats translates into a 50.75 - 49.25 = 1.5 shift in macropartisanship. 10 Our partisanship data were drawn from the NES Cumulative Data File for 1952-1994. The data for each survey were weighted. For surveys with a panel design, only pre-election surveys were used to avoid biases due to panel attrition. 11 In the rst nine polls of 1953, the Gallup organization used the wording, Normally, do you consider yourself a Republican, a Democrat, or an Independent?" 6 argue that the short-term focus of the politics as of today" question accounts for the greater volatility one observes in the Gallup series. This claim rests on their analysis of the biennial National Election Studies. In reply, MacKuen et al. 1992 criticize the NES series on the grounds that it contains too few observations for meaningful time-series analysis and contend that a more informative test of the question wording hypothesis requires a quarterly survey using the SRC wording, such as the CBS New York Times Polls conducted since 1976. In their view, the Gallup and CBS NYT macropartisanship readings seem to track closely over time, despite the fact that the latter employs the SRC question wording.12 The comparison between the Gallup and CBS NYT macropartisanship series is potentially quite informative, but it is complicated by the fact that the Gallup organization surveys many more respondents during the course of a typical quarter. The Gallup series thus contains far less sampling variability. The consequences for the statistical analysis of macropartisan stability can be profound. Measurement error tends to attenuate the autoregressive coe cient obtained when macropartisanship at time t is regressed on macropartisanship at t , 1. When this type of model is used to assess the extent to which current values of partisanship are shaped by past values, the CBS NYT series will be at a disadvantage vis-a-vis the Gallup series. In e ect, sampling error will be misinterpreted as real change in macropartisanship. To draw a meaningful comparison between these two series, therefore, one must use an estimation technique such as two-stage least squares that allows for the possibility of sampling error. Table 1 compares two sets of regressions, one applied to the CBS NYT data; the other, to the Gallup data. Both analyses cover the same time period, from the inception of the CBS NYT poll 1976:1 through the fourth quarter of 1995. First, OLS was applied to the autoregressive model where Ct refers to CBS NYT macropartisanship; Gt , to Gallup: Ct = Gt = 0 + 1 Ct,1 + uCt : + 1 Gt,1 + uGt : 0 If the Gallup series is indeed more volatile, we should expect to see 1 exceed 1 and VARuGt  exceed VARuCt . Ironically, OLS suggests that is greater than : the Gallup series is more strongly shaped by its past values. When two-stage least squares is used to correct for sampling error, the estimated stability of For a parallel debate over the e ects of question wording on the association between party identi cation and short term evaluations in individual level data, see Abramson and Ostrom 1994 and Bishop et al. 1994. 12 7 the Gallup series creeps upward, while the stability of the CBS NYT series grows markedly.13 One obtains similar results when full information maximum likelihood FIML is used to estimate the autoregressive e ects. The estimated autoregressive parameters are .98 CBS NYT and .95 Gallup, suggesting that shocks to macropartisanship take quite a long time to dissipate cf. Box-Ste ensmeier and Smith 1996. We will see momentarily that much of our disagreement with MacKuen et al. turns on the magnitude and origins of these autoregressive e ects. Because FIML enables us to distinguish between sampling variance which is transitory and disturbance variance which continues to in uence subsequent manifestations of partisanship, the analysis also reveals the source of the observed di erences between the two series. Sampling variance amounts to 3.34 percentagepoints in the CBS NYT series but just one percentage-point in the Gallup series.14 On the other hand, the Gallup series contains signi cantly more disturbance variance 2.21 versus .42, p :05, suggesting that it is more strongly a ected by short-term shocks. In sum, the two series are propelled to an equal extent by their past values, but the Gallup series is bu eted more sharply by changes in the environment. The Gallup macropartisanship series used by MacKuen et al., therefore, has its advantages and disadvantages. On the positive side of the ledger is the fact that in comparison to CBS NYT the Gallup series is relatively free from sampling error and extends back to the early 1950s. On the negative side, the short-term focus of the partisanship measure seems to make the series more sensitive to short term disturbances. Since we aim to put the partisan stability thesis to an exacting empirical test, this liability may in fact work to our advantage if we can show that even the Gallup series responds only weakly to short-run uctuations in political and economic circumstances. This is yet another argument on behalf of replicating the MacKuen et al. results. The E ects of Short Term Forces on Macropartisanship What causes these quarter-to-quarter movements in macropartisanship? MacKuen et al. 1989, 1992 argue that the aggregate distribution of party identi cation shifts in response to uctuations in consumer These estimates were obtained from two separate TSLS regression in which macropartisanship at t-2 serves as an instrument for macropartisanship at time t-1. 14 Granger causality tests indicate strongly that the Gallup series leads the CBS NYT series by a quarter, but not vice versa. It is tempting to infer that the as of today" wording picks up changing party attachments more quickly, and that general partisan attachments align themselves with one's present-day feelings. A methodological explanation is that on account of their smaller quarterly samples, CBS NYT data contain more measurement error than the Gallup data. As a result, the Gallup polls depict turning points more accurately than the CBS NYT polls. 13 8 con dence and presidential approval. The more popular a sitting president, the more attractive is his party as an object of identi cation. This hypothesis can be tested by tracking inter- or intra-administration changes in presidential approval ratings e.g., the transition from a disgraced Nixon in 1974 to a more popular Gerald Ford, or from the triumphant Reagan of 1984 to the scandal-beset Reagan of 1987. In addition, MacKuen et al. argue, parties are rewarded or punished for their economic stewardship. The more buoyant the public's view of the economy, the more attractive the incumbent president's party. The e ects of the economy may be mediated to some extent by presidential approval, but one way or the other partisanship should follow the business cycle, with the party in power bene tting from an atmosphere of economic optimism.15 MacKuen et al. 1989 examine macropartisanship, presidential approval, and consumer sentiment on a quarterly basis for the period 1953-1988. Approval ratings are compiled from Gallup polls, aggregated quarterly. Economic perceptions are measured using the Michigan Index of Consumer Sentiment MICS.16 The economic and political volatility of this time span is apparent from Figure 2. As one might expect, economic expectations fell during the recessions of late 1950s, mid 1970s and early 1980s. These economic setbacks seemed to have damaged presidential popularity, as did scandals and international problems. Yet, in relation to the roller coaster patterns in economic performance and presidential popularity, the time-series of macropartisanship seems rather placid, even when it seems to move in tandem with these short-term forces. The precipitous drop in Nixon's popularity following his landslide election victory in 1972 coincides with a gradual increase in Democratic identi cation; similarly, Carter's fall from public favor in 1977-8 led to only modest gains for the Republicans. Granted, Reagan's 1980 victory coincides with a resurgence in Republican identi cation, and further gains were registered after the end of the 1982 recession. But to the naked eye, macropartisanship seems fairly stable in the wake of changing political fortunes. We now examine whether this ocular regression is con rmed by more systematic statistical analysis. Replication of the MacKuen et al. 1989 Study MacKuen et al. 1989 employ three statistical tests to establish that partisanship changes in the wake of economic conditions and presidential popularity. First, they conduct a sequence of Granger-Sims exogeneity tests Table 1, p.1132, the purpose of which is to show that change in the Gallup series is anticipated by previous changes in approval and consumer sentiment. Next, a rather elaborate Box-Jenkins time-series model is developed and estimated Tables 2 and 3, p.1135-6 in which approval and economic perceptions 15 In contrast to Weisberg and Smith 1991, MacKuen et al. use consumer perceptions rather than objective economic indicators, such as in ation or unemployment rates. 16 For a description of the MICS survey questions and their relationship to objective economic indicators, see Matsusaka and Sbordone 1995. 9 produce immediate as well as delayed e ects on macropartisanship. Third, the analysis is repeated using ordinary least squares regression Table N-1, pp.1140-1 in an e ort to assure readers that these ndings are not reducible to mere technical wizardry" p.1141. Again, partisanship is found to change in response to swings in approval and consumer con dence. Exogeneity Tests. As a preliminary exercise designed to lend credence to the idea that partisanship responds to short-term changes in the political and economic environment, the authors perform a set of exogeneity tests on three series, macropartisanship, consumer sentiment, and presidential approval. This statistical procedure examines each pair of variables for a total of six tests to see whether a given series is a leading indicator of another series. The procedure p.1132 involves i prewhitening each series by modelling it as a function of its previous values and ii modelling each prewhitened series as a rst order transfer function of a lag-1 whitened independent variable, making allowances for the fact that the sign of the e ects of approval and consumer sentiment on macropartisanship will depend on which party is in o ce.17 MacKuen et al. seek to show that lagged presidential approval and lagged consumer sentiment predict macropartisanship, but not the reverse. The results they report p.1132 suggest a strong, unidirectional causal ow from approval and consumer sentiment to partisanship. Our results, using the complete Gallup series, are inconclusive. As shown in Table 2, consumer sentiment seems to have no e ect on the Gallup series. Approval's in uence is stronger, but judging from the test in which partisanship is a predictor of approval, the causal ow could be said to run in the opposite direction. Note that these discrepancies turn up primarily in tests involving the updated Gallup series. The two tests using the approval and consumer sentiment series produce results similar those reported by MacKuen et al. As these authors note, changes in consumer sentiment anticipate changes in presidential approval and not the reverse. No comparable pattern of clear temporal sequencing emerges from the analysis of partisanship. The crosscorrelations between party identi cation and each of the other series also provide equivocal diagnostic results. Table 3 presents the crosscorrelations to lag-12 for the prewhitened series generated above, for series generated by more extensive prewhitening procedures, and for series created merely by taking the rst di erence of each variable.18 As it turns out, each whitening technique generates similar results, since the Taking these sign changes into account makes the analysis of macropartisanship in some ways more complex than models of presidential approval. As we note below, sign changes can be addressed in a number of di erent ways, but all generate results similar to those reported in this section. 18 The more extensive prewhitening procedure for approval involved including dummy variable for the rst quarter of each new administration in addition to an 1,0,1 error model. For MICS, prewhitening required an AR1, MA2, MA3, and MA4 speci cation. 17 10 three series are all strongly autoregressive. Again, the resulting MICS and approval scores during Republican administrations were multiplied by -1 so that the anticipated correlations with the macropartisanship series would be positive. Lagged consumer sentiment bears an erratic and theoretically unexpected relationship to the macropartisanship series. Regardless of which set of crosscorrelations are consulted, we nd no contemporaneous correlation nor any lag-1 relationship. Instead, we see two spikes both of marginal statistical signi cance at lags 2 and 6, with a valley in between. The crosscorrelation functions linking approval and macropartisanship look more sensible but do not furnish strong evidence of approval's predictive power. In each case, the contemporaneous correlation is a statistically signi cant .26 or .27, but the lag-1 correlation drops just below conventional signi cance levels. At higher lags, the crosscorrelations are negligible. The contemporaneous correlations are potentially suspect, however, since the Gallup approval scores are obtained from the same surveys as the macropartisanship measures; sampling errors will therefore tend to be correlated. Quarters that sample a disproportionate number of Republicans will tend also to have higher approval scores for Republican presidents. The Granger-Sims test used by MacKuen et al. guards against this problem by lagging each predictor series as does MacKuen et al.'s 1989 and 1992 OLS analysis; see below, but the Box-Jenkins model described in the next section allows approval's e ects to diminish geometrically with time beginning at lag-0. Box-Jenkins Time Series Model. The key statistical analysis presented by MacKuen et al. 1989 is a Box-Jenkins ARMA model. The model is estimated in two stages. First, approval is modelled as a function of consumer sentiment, historical events, and administration e ects. This preliminary step is taken to purge approval of the variance attributable to consumer sentiment.19 Second, partisanship is modelled as a function of consumer sentiment, purged presidential approval, and a set of historical events and administration e ects. Each administration, for example, is represented by both a dummy variable and a template that begins with a score of one at the president's initial inauguration and then declines exponentially throughout his tenure in o ce" by a prespeci ed rate p.1133. The dummy and template variables help smooth transitions across Democratic and Republican administrations, but dummy variables for time periods have the practical consequence of undercutting the apparent autoregressive character of macropartisanship McCleary and Hay 1980. Had we inserted administration dummies into the analysis presented in Table 1, the e ects of lagged macropartisanship would have declined by about 40. Note that the two equation approach is designed to credit consumer con dence for the in uence it exerts 19 This variable is approval minus that part of approval forecasted from the economic component alone, with other parts of the model zeroed out" p.1140, note 10. 11 indirectly through presidential approval. In other words, the model is designed to estimate the total e ect of consumer sentiment on macropartisanship. Any speci cation of this sort immediately raises questions about the robustness of the ndings. How do the results change, if at all, when approval or consumer sentiment are entered into the model in di erent ways? Do we obtain a similar assessment of consumer sentiment's total e ect when we omit presidential approval from the model? Another important modelling nuance is the use of dynamic" parameters  k  in the Box-Jenkins model. These parameters describe the rate at which changes in the political environment that occur in one period decay over subsequent periods. MacKuen et al., in other words, express macropartisanship Mt  as a decaying function of approval At  and consumer sentiment Ct , a matrix of administration and event variables Z , and an AR1 disturbance ut : u Mt = 1 , 1 L At + 1 , 2 L Ct + Z + 1 , t L ; 1 2 3 where L represents the backshift operator, such that LAt = At,1 . A dynamic parameter value of zero implies that an independent variable's in uence persists for just one quarter. Conversely, a value of one means that an independent variable's impact is felt each and every subsequent period, without decaying over time. In between these extremes, a parameter of .5 implies that an e ect is full for the rst period, halved for the next, halved again, and so on. When an independent variable undergoes a permanent one-unit shift, the cumulative long-term e ect given an immediate impact of and a dynamic parameter would be =1 , . One of the central ndings of MacKuen et al. 1989 is that the three estimated dynamic parameters vary markedly. MacKuen et al. obtain a large dynamic parameter estimate for consumer sentiment .84, suggesting that changes in economic perceptions continue to in uence macropartisanship long after they occur p.1137. Approval seems to have less staying power .35, but there is again clear evidence of gradually dissipating in uence over time. On the other hand, the AR1 coe cient  3  takes on the surprising value of -.04, suggesting that consumer sentiment, political approval, and the assorted administration variables capture all of the dynamics in the macropartisanship series seen earlier in Table 1. Our replication begins with the Box-Jenkins model of presidential approval. With the exception of certain administration e ects, most coe cients are similar to those reported by MacKuen et al. Particularly close are the estimated coe cients for the immediate impact and dynamic parameter associated with consumer sentiment. Table 4 indicates that a one percentage-point shift in consumer sentiment results in an immediate .33 change in presidential approval, followed by e ects that decline by a factor of .64 thereafter. 12 Comparable estimates obtained by MacKuen et al. are .32 and .61, respectively.20 Given the closeness of the results, we are reasonably con dent that our model replicates the original analysis, the minor discrepancies being attributable to post-publication adjustments to the approval and MICS series by MacKuen et al. Following the procedure of purging approval using the parameter estimates obtained by MacKuen et al., we attempted to reproduce the results reported in Table 3 of MacKuen et al. 1989. We have included all of the regressors that MacKuen et al. include in their Table 2, not just the statistically signi cant predictors that MacKuen et al. list in their Table 3. Replication of the analysis with just the subset of regressors listed by MacKuen et al. in their Table 3 drives all of the estimated short-term e ects to nil. We have also included dummy variables to mark the rst quarter during which a new president was in o ce. In e ect, these variables eliminate from consideration those observations in which lagged approval would refer to a di erent incumbent president. It is important to stress that although these transition dummies would seem redundant, without them, one obtains no evidence of short-term in uences to speak of. What we present here are the maximal estimates of short-term e ects that we uncovered during our e orts to replicate the MacKuen et al. model. Table 5 compares our results to those obtained by MacKuen et al. In Replication 1, we come fairly close. The dynamic coe cients are approximately the same, suggesting that the e ects of a shock in the consumer sentiment series stretch over years. The decay in the e ects of approval occurs more rapidly, as MacKuen et al. suggest. The principal di erence between our results and those obtained by MacKuen et al. lies in the magnitude of the contemporaneous e ects. Our coe cient for approval is half of what was previously estimated .117 versus .22; for consumer sentiment, two-thirds .066 versus .10. Another di erence is that we nd a stronger AR1 coe cient, although this estimate is still much lower than what was observed in Table 1 because MacKuen et al.'s model includes dummy variables for each administration. Finally, and not without important implications for the argument that partisanship wavers amid shifting party fortunes, we nd no signi cant direct e ects of historical events on macropartisanship. In order to place the Box-Jenkins results into the context of other statistical analyses undertaken by MacKuen et al. 1989, 1992, we experimented with alternative model speci cations. Replication 2 omits approval from the model but obtains similar estimated e ects for consumer sentiment, con rming that the original MacKuen et al. model estimates consumer sentiment's total e ect on macropartisanship. 20 The most noteworthy discrepancy between the two sets of results is the larger AR1 term we obtain. Our analysis, in other words, suggests that presidential popularity is more self-perpetuating. We interpret the discrepant signs of the administration coe cients to di erent ways in which these variables have been scored for certain administrations. 13 Replication 3, which substitutes presidential approval for political" approval, establishes that consumer sentiment exerts no direct e ect on macropartisanship, net of presidential evaluations. Replication 4 discards consumer sentiment and focuses on presidential approval which no longer had to be purged, since consumer sentiment was absent from the model. We nd statistically signi cant coe cients for the immediate impact of approval and its rate of decay, but both are notably smaller than what MacKuen et al. report. For example, MacKuen et al.'s results imply that a sustained one-unit shift in the approval series leads to a cumulative shift in macropartisanship of .34, whereas our results place the estimate at .20. Finally, anticipating the OLS analysis that follows, we estimated the original MacKuen et al. model, lagging both consumer sentiment and political approval one period. Replication 5 fails to nd signi cant immediate or dynamic parameter estimates for either variable p :05. Ordinary Least Squares Regression. MacKuen et al. present a variant of this Box-Jenkins model in the endnotes to their essay p.1140-1. Regression is applied to a model that proceeds in the same two-step manner. First, approval is regressed on the same administration and events variables, approval lagged one quarter, and consumer sentiment lagged one quarter. The lagged e ect of consumer sentiment .29, see Table N-1 is then subtracted from approval to create a purged approval series. The second step equation modelling macropartisanship includes the lag-1 purged approval series, lag-1 consumer sentiment, lag-1 macropartisanship, and the rest of the event and administration variables p.1140, note 13.. Although this model would seem quite di erent from the Box-Jenkins model discussed earlier, it can be rewritten in a form that is roughly comparable to Replication 5 in Table 5. Again, let Mt represent macropartisanship, and let At,1 denote lagged political approval, and Ct,1 lagged consumer sentiment. Call Z a matrix of administration and event variables and ut a white noise disturbance: Mt = Mt,1 + 1 At,1 + 2 Ct,1 + Z + ut Mt , Mt,1 = 1 At,1 + 2 Ct,1 + Z + ut 1 , LMt = 1 At,1 + 2 Ct,1 + Z + ut 1 u Mt = 1 ,1 L At,1 + 1 ,2 L Ct,1 + 1 , L Z  + 1 , t L : By comparison, the model underlying Replication 5 in Table 5 is: u Mt = 1 , 1 L At,1 + 1 , 2 L Ct,1 + Z + 1 , t L ; 1 2 3 Thus, the OLS model and the Box-Jenkins di er according to the number of dynamic parameters  k  and the speci cation of the transfer function linking Z to Mt . The models converge in the special case in which 14 =0 an assumption that will be invoked in the revised model we present below and where the three k parameters in the Box-Jenkins model are equal. Note that the latter constraint is inconsistent with MacKuen et al.'s earlier nding that consumer sentiment's e ects decay more slowly than approval's. MacKuen et al. present statistically signi cant, although not substantively large, e ects for both purged approval .13 and consumer sentiment .10. The decay parameter the coe cient for lagged partisanship is estimated at .26. The authors seem to regard these results as con rming their thesis about partisan changability, but in fact these estimates suggest the limited in uence of economic and political perceptions. MacKuen et al.'s Box-Jenkins analysis indicates that a permanent one-point shock in consumer sentiment eventually produces a .63 adjustment in macropartisanship; their OLS results imply an adjustment of .14. Our e orts to replicate this regression analysis proved successful for the model in which presidential approval is the dependent variable. As shown in Table 6, the lagged e ects of consumer sentiment and approval are .37 versus .29 and .38 versus .31, respectively, and both are highly signi cant. Our R-square is .93; theirs, .94. Not so, when the dependent variable becomes the comprehensive Gallup macropartisanship series. Each of the speci cations we tried produced results that were weaker than those reported by MacKuen et al. The e ect of political approval is small and statistically insigni cant. Replacing political approval with presidential approval increases the estimate, but it remains half the size of that obtained by MacKuen et al. The e ect of consumer sentiment seems more robust, particularly when approval is omitted from the regression equation. Because the dynamic coe cient .53 is found to be larger in our analysis, the cumulative impact of a sustained one-point shock to popularity or consumer sentiment is similar to that implied by MacKuen et al.'s results which is to say relatively small. To summarize, the three analyses of the Gallup data consistently turn up weaker short-run adjustment e ects than those reported by MacKuen et al. Cross-Validation Using CBS NYT Polls In an e ort to disprove Abramson and Ostrom's claim that the Gallup series moves with short-term changes in the political environment due to the immediate focus of its party identi cation measure, MacKuen et al. 1992 replicate their results using CBS NYT polls, which use the SRC question wording. These polls extend back to 1976, but MacKuen et al. 1992 analyze only the 32 quarters of the Reagan administration. Apparently due to the paucity of observations, the authors rely exclusively on an OLS regression analysis and restrict the model of macropartisanship to three lagged regressors: macropartisanship, consumer sentiment, and presidential approval. Approval is used in its original form; no purging procedure is used. 15 Table 7 presents the comparison between our results and those obtained by MacKuen et al. 1992.21 Our results using the CBS NYT polls are similar, our analysis producing the same e ect for approval and a slightly weaker e ect for consumer sentiment. The Gallup results di er a bit more, with our model generating a higher R2 and suggesting a weaker e ect of presidential approval. In short, our results are similar but not identical to those reported by MacKuen et al. 1992. These results corroborate the OLS analysis undertaken in MacKuen et al. 1989 in that the estimated e ects of short-term forces are quite modest. And indeed the estimates become more modest still when the analysis is replicated over a wider time horizon. CBS NYT polls date back to the rst quarter of 1976, and we have updated both the Gallup and CBS NYT series into the Clinton administration. As Table 7 makes clear, the addition of some three dozen new cases reduces the apparent e ects of presidential approval and consumer sentiment. Of the two, only economic perceptions have a statistically reliable, though small, immediate impact .056, .050 on macropartisanship. In sum, the CBS NYT series corroborates the results obtained using Gallup data; both yield minimal support for the notion that partisanship shifts abruptly in response to changes in political and economic perceptions. A Revised Model and Out-of-Sample Forecasting Tests To this point, we have worked within the modelling framework laid out by MacKuen, Erikson, and Stimson in an e ort to assess the robustness of their results. In this section, we present a revised model that preserves the essential logic of previous analyses but is more exible and parsimonious. A further advantage of the revised model is that it greatly simpli es the task of producing out of sample forecasts, an exercise that enables us to evaluate the claim that partisanship is bu eted by changing economic and political evaluations. One of the technical challenges one confronts when modelling macropartisanship is that the signs of the key coe cients change depending on which party is in o ce. Good economic times and favorable approval ratings add to the stock of Democrats during a Democratic administration and subtract during periods of 21 Producing an exact replication of the results reported by MacKuen et al. 1992 is hampered by the fact that they report an analysis using just 26 of the 32 cases. The authors restricted their analysis to cases in which both Gallup and CBS-Times measures of macropartisanship are measured for times t and t , 1" p.479. The rst quarter's observation is lost when the data are lagged one period. Another two cases are lost because no CBS NYT polls were conducted during the third quarter of 1981. This leaves 29 cases. The authors do not tell us which of the remaining quarters were omitted. We tried out every possible combination of 26 cases but were unable to replicate the published results. We came closest to the reported CBS NYT results upon dropping 1982:4 and 1988:4, but the Gallup results remained more or less as we report them. 16 Republican control. MacKuen et al. address this problem by multiplying consumer sentiment and presidential approval by -1 during Republican administrations, but complications mount when macropartisanship is modelled as a decaying function of lagged approval or consumer sentiment, as in the Box-Jenkins model discussed above. Consider a simpli ed example in which Mt = 40 + 1 , L At ; where = :25 and = :5. Suppose that during the sixteen quarters of a Democratic administration, approval At  stands at 50; thereafter, a Republican president assumes power and also receives an approval rating of 50 during each of the rst four quarters in o ce. Mt stands at 65 during the last quarter of the Democrat's term. Because the model makes no allowance for party control of the Presidency, it predicts that Mt will drop to 52.5 during the rst quarter of the new administration and continues to diminish geometrically until it reaches 40. This anticipated shift in party identi cation is counterintuitive, since the approval scores for both presidents are exactly the same. MacKuen et al. attempt to eliminate these artifactual changes in party identi cation by adding administration dummies and variables marking the transition between administrations. Our alternative approach involves generating di erent approval and consumer sentiment series for each president and allowing the signs of the coe cients to link approval and partisanship in an appropriate manner across administrations. For example, Jimmy Carter's approval score is set to zero during quarters before 1977 and after 1980, and to his actual approval score minus its 1977-1980 mean during his quarters in o ce. Following a similar procedure for the other presidents produces a sequence of approval and consumer sentiment variables. One may then place constraints on the parameters associated with these variables to test the sequences of hypotheses that i all approval variables for presidents of a given party have the same e ect on macropartisanship and ii that the coe cient associated with Democratic presidents' approval is -1 times the coe cient for Republican presidents. This approach obviates the need for transition templates and the stipulated decay parameter that goes with them, administration dummies, and start-of-term dummies. At the same time, the model has the advantage of testing the stability of approval and consumer sentiment e ects over time.22 Table 8 lays out a sequence of nested models linking consumer sentiment and or approval to macropar22 A conceptual issue lurks beneath these arcane details. This model does not assume comparability of approval and consumer sentiment series over time. Arguably, the criteria used to evaluate economic and presidential performance|and hence the meaning of particular numerical ratings|change over time. Our model allows a test of this proposition. 17 tisanship for the period 1953:1 1988:4. The rst row of Table 8 presents the null model, in which approval and consumer sentiment variables are omitted entirely. The second row introduces consumer sentiment and approval using the most restrictive speci cation, one that constrains each of the approval or consumer sentiment parameters to be equal in absolute value. The third row relaxes the constraint that Democratic and Republican e ects are equal in size and opposite sign. The fourth row allows short-term forces during each administration to have varying e ects. Wald tests for coe cient restrictions clearly favor the Constant E ects Model. The reduction in residual sum of squares does not allow one to reject the null hypothesis that short-term forces have the same in uence on partisanship during periods of Republican and Democratic incumbency. Nor do the data support the hypothesis that short-term forces have distinctive in uences during di erent presidencies. The results, in other words, suggest that macropartisanship may be modelled adequately by a very parsimonious model featuring a single approval parameter and a single consumer sentiment parameter. This nding will doubtless come as welcome news to analysts of macropartisanship data, whose work is made easier by simpli cations of this kind. The estimates obtained from this constrained model rea rm the conclusion that macropartisanship adjusts to short-term shocks in a limited and gradual fashion. In Table 9 we see that CBS NYT data o er uniformly weak support for the hypothesis that macropartisanship travels with presidential approval or consumer sentiment. Using the Gallup series, the strongest e ects of short-term forces are obtained when partisanship is modelled as a function of contemporaneous, rather than lagged, values of approval and consumer sentiment. As noted earlier, this contemporaneous speci cation potentially con ates sampling covariation and real movement in partisanship due to shifts in approval; it is telling that contemporaneous Gallup approval has a smaller impact on the CBS NYT macropartisanship series. Even so, the Gallup results suggest that a 21 percentage-point shock in contemporaneous presidential approval is required to produce an immediate adjustment in macropartisanship of just one percentage-point. When lagged approval and lagged consumer sentiment are used as predictors in a manner approximating the OLS speci cation used by MacKuen et al., weak and statistically insigni cant coe cients turn up, albeit of proper sign. Since MacKuen et al. 1989 emphasize the distinctive dynamics of consumer sentiment and presidential approval, we conducted a series of additional analyses that allow for varying dynamic parameters. Using both Gallup and CBS NYT data, we examined the reduction in error variance achieved by relaxing the constraint that 1 = 2 = 3 in the model u Mt = 1 , 1 L At + 1 , 2 L Ct + 1 , t L : 1 2 3 In neither case does a Wald test enable one to reject the null hypothesis that a single dynamic parameter 18 generated the data p :10. In one sense, we have come full circle. The Box-Jenkins analysis presented by MacKuen et al. suggests that the macropartisanship series has no autocorrelation once one controls for the dynamic e ects of short-term forces. Our ndings, by contrast, emphasize self-replicating character of partisanship and the uniform dynamics underlying the macropartisanship series. Our primary disagreement, however, hinges on the magnitude and robustness of the immediate shortterm e ects. One way to illustrate the elusiveness of these e ects is to perform an out-of-sample forecasting experiment. If MacKuen et al. are right about the in uence of consumer sentiment and presidential approval, models that include approval or consumer sentiment should outperform models that ignore these predictors. We evaluated the predictive accuracy of the competing models by forecasting the 28 quarterly observations from 1989:1 through 1995:4 using the parameter estimates obtained from our analyses of 1953:1 through 1988:4. Forecasts were generated dynamically, which is to say that values of the dependent variable at time t were used to forecast outcomes at t + 1. Whether gauged by the criterion of mean absolute error or root mean squared error RMSE, the forecasting accuracy of the null model is on par with models incorporating approval and or consumer sentiment. Returning to Table 8, we see, for example, that the RMSE of the null model is 2.09. When contemporaneous consumer sentiment is added as a predictor, the RMSE remains at 2.07. Only when one adds contemporaneous approval does the RMSE drop to 1.98, but this modest improvement is to some degree an artifact of sampling covariation. Replacing these predictors with their lagged values in a manner akin to MacKuen et al.'s 1992 speci cation slightly reduces forecasting accuracy. Nor does predictive accuracy improve when one experiments with di erent lags, such as the phantom lag-2 or lag-6 relationships that turned up in crosscorrelations discussed earlier. Much the same pattern of forecasting results obtains when the CBS NYT macropartisanship series is used in place of the Gallup series.23 Again, only contemporaneous presidential approval contributes to predictive accuracy, and marginally at that. Discussion Findings purporting to show that partisanship travels with short term forces have had a demonstrable impact on the way students of electoral politics have come to view party identi cation. Tracing successive editions of leading texts such as the Change and Continuity series Abramson, Aldrich, and Rohde 1982: 161, 1995: 247 or Controversies in American Voting Behavior Neimi and Weisberg 1976: 310, 1993: 268, one nds increasing skepticism toward the traditional view of party identi cation as an unmoved mover. Like MacKuen, Erikson, and Stimson, many students of public opinion now see in decades of survey data clear evidence that partisanship's twisting course has been shaped by the winds of political and economic 23 These results have been omitted from Table 8 for the sake of brevity but are available on request. 19 fortune" p.1139. Although we take a di erent view of macropartisan change, this metaphor nonetheless seems apt. Wind, unlike water, alters a landscape slowly, almost imperceptibly. A political party is doubtless more successful at winning over adherents when its leaders are popular and its record spotless. Skeptical though we may be of the notion that short-term changes in the political environment alter the distribution of party identi cations appreciably, it would be foolish to take the position that party attachments are altogether unwavering. The question is whether shifts in partisanship are large enough to call into question traditional views of realignment and the stabilizing role that party identi cation plays in a party system. MacKuen, Erikson, and Stimson argue that the relationship between short-term uctuations and partisanship is so strong that one must speak of realignment as an ongoing process, one in which the party balance at any given point in time re ects recent trends in political popularity and economic optimism. Our results call this characterization into question. The sluggish pace at which partisanship adapts to changing circumstances means that only a dramatic and sustained shift in the political fortunes of a party can precipitate sizable changes in the macropartisan balance. Given the ebb and ow of partisan warfare and economic performance in the United States, it is understandable that substantial, let alone cataclysmic, partisan realignments should turn up so infrequently. Lacking the exact dataset used by MacKuen et al., we cannot say precisely why we come to divergent conclusions. But having reproduced their analyses using improved measures, extended their analysis to a wider timespan, and checked the results against those obtained from a reformulated model, we conclude that partisanship responds to the political environment in a limited and gradual fashion and that the tension between micropartisan stability and macropartisan instability is more apparent than real. The broader methodological lesson here concerns the importance of returning to pivotal empirical demonstrations. We have by no means exhausted the supply of statistical tests or alternative speci cations that might be applied to the macropartisanship data, and our conclusions may one day be overturned by more cogent and telling analyses. But in the end the general quality of research in this area will have been improved by wider participation in the research process. 20