We address two questions: How do -people form their expectations about the likely winner of

Transcription

1 American Political Science Review Vol. 88, No. 2 June 1994 EXPECTATIONS AND PREFERENCES IN BRITISH GENERAL ELECTIONS RICHARD NADEAU Universite de Montreal RICHARD G. NIEMI University of Rochester TIMOTHY AMATO University of Iowa We address two questions: How do -people form their expectations about the likely winner of the next general election? and What are the links between expectations and votes? Using data collected by the Gallup organization in Great Britain, we find that the expectations formation process (1) has a significant inertia component but also a rapid adjustment to current information; (2) reflects voters' ability to translate economic expectations into political forecasts; and (3) is "time-bounded," possessing special characteristics immediately before and after a general election. The analysis also confirms the existence of a small bandwagon effect, whereby expectations that one party will win inflate that party's vote. The ability of voters to make reasonable forecasts without being unduly influenced by their own preferences suggests that under normal circumstances voters are expressing real preferences and not simply following the crowd. T he idea that observed contemporary behaviors are strongly linked to expectations about future outcomes has become so central for economists that much of economic theory has been rewritten to take explicit account of this concept (Begg 1982; Hudson 1982, chap. 3). Political scientists have also shown a growing interest in expectations, with future prospects proving to be particularly useful in studying certain aspects of electoral behavior. The now-common use of the notions of viability and electability, as well as the classification of voters as straightforward, sincere, or strategic, exemplify this trend (Abramson et al. 1992; Bartels 1988; Niemi, Whitten, and Franklin 1992). The simple act of voting is increasingly conceived of in terms of utilities (voters' preferences) and probabilities (voters' assessments of candidates' chances of winning). Two crucial questions are at the nexus of the votes-expectations literature: How do people form their expectations and What are the links between expectations and votes? We address these questions with data collected by the Gallup organization. This unique data set includes not only the usual measurement of voters' preferences among parties but also a systematic tracking of voters' expectations about the outcome of the next general election. 1 The fact that this information is available on a monthly basis for more than a decade permits systematic study of the expectations formation process and of the links between expectations and votes not only during election periods but also in less intense, interelection times. 2 The data are thus attractive despite their obvious limitations. 3 We shall present a simultaneous-equation votesexpectations model, 4 paying particular attention to the expectations formation process. 5 We then present our findings, which show a modest effect of expectations on voting behavior and stronger and more variegated effects including "instantaneous" adjustments of voting intentions on expectations. We discuss the theoretical implications of our findings in the conclusion. A VOTES-EXPECTATIONS MODEL The votes-expectations literature has been mainly concerned with the existence and form of a pattern of reciprocal causation between voters' preferences and their assessments of the electoral campaigns (i.e., the parties' standings). Consistent with this perspective, we examine here two possibilities: (1) that voting intentions are affected by expectations about the likely winner of the next general election (bandwagon effect) and (2) that expectations about which party will win the next election are dependent on current vote intentions (whether through a sophisticated treatment of information or a simple projection effect). Practically speaking, what is required is that we devise and estimate a simultaneous equations system formed of a vote equation and an expectations equation. As is standard in simultaneous equations models, we estimate our model using two-stage least squares. To obtain consistent estimates, we must pay particular attention to model specification and estimation, especially the less-frequently-studied expectations equation. Substantive questions concerning the appropriate specification of the model will be discussed. Identification issues, estimation strategy, and diagnostic tests are described in Appendices A andb. The Vote Equation To specify the vote equation of our simultaneousequation model, we build from the extensive literature on the short-term dynamics of political party popularity in Britain (Alt 1979; Clarke, Mishler, and Whiteley 1990; Clarke, Stewart, and Zuk 1986; Dunleavey and Husbands 1985; Frey and Schneider 1978; 371

2 Expectations in British Elections June 1994 Goodhart and Bhansali 1970; Hibbs 1982; Hudson 1985; Lewis-Beck 1988; Mishler, Hoskin, and Fitzgerald 1989; Miller and Mackie 1973; Norpoth 1991; Pissarides 1980; Sanders, Ward, and Marsh 1991). In particular, we searched for a satisfactory mix of economic and political indicators to serve as control variables whose inclusion would prevent us from inflating the impact of the expectations term. The specification we arrived at follows closely the prescription of Clarke, Mishler, and Whiteley: A properly specified model of government popularity requires four broad classes of explanatory variables:... the long-term predispositions of voters,... issue perceptions which include retrospective and prospective thinking about economic and noneconomic issues,... leadership images, which relate both to the government and opposition parties,... [and] various systematic and random shocks to the system. (1990, 66) Our specification is shown in equation 1: VCL, = j8 0 + _! + /3 2 MPSHARE f + /3 3 UNEMP, _ j6 6 LEADERS( )8 8 WCL f, (1) where VCL is the percent intending to vote Conservative minus percent intending to vote Labour (percent calculated after excluding don't knows, undecideds, refusals); MPSHARE is the ratio of the percent intending to vote for a minor party to the sum of the percents intending to vote Conservative or Labour; UNEMP is the unemployment rate, monthly, seasonally adjusted; INF is the monthly inflation rate calculated from the adjusted retail price index; PEXP is the percent thinking that personal economic conditions will improve over the next 12 months minus percent thinking that personal economic conditions will worsen over the next 12 months; LEADERS is the percent approval of the prime minister minus percent approval of the opposition leader, FALK is a dummy variable equal to 1 for May-June 1982 (the Falklands War), 0 otherwise; and WCL is the percent thinking that the Conservatives are most likely to win the next general election minus percent thinking that Labour is most likely to win. 6 The coding of the variables and the form of the equation deserve brief explanation. Our major concern, of course, is with the impact of expectations about the winner WCL. 7 It is assumed that this impact is instantaneous, that is, that the gap measured in the current poll is already reflected in vote intentions. The assumption of immediate effects will be discussed in considerable detail, when we make a similar assumption about the effects of vote intentions on expectations. The expectations variable itself is expressed as a percentage difference because the popularity gap between the parties is at the core of thinking of one as a potential winner and the other as a loser. Likewise, the vote intentions variable is measured as a percentage difference. The most interesting of the remaining variables in the vote equation is VCL t _]. Under normal circumstances vote intentions "do not move abruptly from j one moment to the next one, but at a sluggish pace instead" (Norpoth 1991,146). We include the lagged dependent variable to capture this inertia component. Following Norpoth, this variable may be conceived as reflecting a "short-term memory" process, whereas our constant term may be an indicator of the "long-term component of popularity" (pp ). Based on Mishler, Hoskin, and Fitzgerald's (1989) results, we also checked, by including the MPSHARE variable, whether the party balance between Conservatives and Labour was affected by the electoral fortunes of third parties. As theory and recent empirical work suggest that voters' evaluations of government economic performance may be a blend of both objective conditions and subjective perceptions, 8 we added to our vote equation the unemployment rate, the inflation rate, and a variable reflecting voters' prospective assessments of their personal economic situations over the next 12 months. 9 The inclusion of unemployment and inflation appears sensible because both are "highlighted by the media, and salient to the voter" (Clarke, Mishler, and Whiteley 1990,66; see also Mosley 1984). 10 The subjective perception variable (PEXP) was selected after various tries with similar variables. 11 Inclusion of the variable LEADERS follows what has become standard practice in British studies. Given the finding by Mishler, Hoskin, and Fitzgerald that "the relationship between leader satisfaction and party support is highly significant" (1989, 228), it is also consistent with our goal of not inflating the expectations term in our vote equation. Moreover, this inclusion also seems justified on theoretical grounds, as "recent research suggests that although related, leadership satisfaction and party popularity are conceptually distinct and empirically separable (Clarke, Mishler, and Whiteley 1990, 68). 12 Finally, the Falklands War exemplifies a major disruption to the usual pattern of stability, and a variable was included to take it into account. 13 The Expectations Equation In the search for an appropriate specification for our expectations equation, we faced more difficulties than for the vote equation. First, while popularity functions have long been available and have been extensively studied, such is not the case for expectations; there is no extensive body of research on expectations comparable to that for vote intentions. Second, our dependent variable the expectations question lacks the immediate time horizon of the vote question (who will win "the next general election" versus one's vote if there were a general election "tomorrow"). Some insight can be derived from work on expectations in economics. Early attempts to model the expectation formation process centered on the "adaptive" expectations model, where lagged values of a variable were used to predict current inflation or other relevant economic quantities (e.g., Cagan 1956 and Nerlove 1958; a good summary is Nerlove, Gre- 372

3 American Political Science Review Vol. 88, No. 2 ther, and Carvalho 1979, chap. 10). These initial attempts were subsequently criticized as being "atheoretical" and inconsistent with any explicit optimizing behavior on the part of individuals within the model. Later research in economics argued that expectations were formed "rationally" or in a "model consistent" manner (Lucas 1973; Nerlove, Grether, and Carvalho 1979, chap. 10; Sargent and Wallace 1975); that is, expectations of an economic variable should be based upon all the variables in the model and not just the past values of that variable. 14 Our approach to modeling the expectations formation process lies somewhere between these two polar cases. Like the early critics of the "adaptive" expectations process, we recognize that more than lagged values of a variable are required to explain how individuals make forecasts of quantities they find important. However, we also recognize that the rational expectations model used in economics may be too strong for our explicitly political analysis. In particular, an empirical economic model that incorporates the rational expectations assumption will typically include a series of economic variables in the expectations equation. This is because economic agents are assumed to use these variables in forecasting, say, inflation or unemployment. Given the time horizon of the economists' expectations model, this strategy makes perfect sense. One is typically interested in forecasting inflation or any other economic variable for the next period. In our political model the argument is different. The agents in our model are voters, and they are often asked to forecast an event (the election outcome) many periods from now. Consequently, economic information in the form of inflation and unemployment rates will not be particularly useful since current values provide very little predictive power for economic variables possibly two or three years from today. 15 Other restrictions to the expectations equation appear plausible as well. The leader variable used in the vote equation is strongly retrospective rather than prospective, 16 and the minor party share variable certainly does not capture the complexities related to the translation of small parties' short-term success into major parties' future electoral prospects. Finally, we theorize that single events such as the Falklands War are unlikely to influence expectations except insofar as they leave their imprint on the parties' standings. The variables that will be important in forming our "political" expectations are those with an avowedly political component, including such factors as parties' levels of support, party reputation, and the period in the electoral cycle, plus one economic variable that asks respondents to look forward a full year. The following equation reflects this mode of thinking: 17 WCL ( = a 0 + «iwcl t _! + a^ostelect * WCL ( _ i a 3 VCL, + a 4 PREELECT * VCL f a 6 D8387, where WCL, VCL, and PEXP are defined as in equation 1; POSTELECT is a dummy variable equal to 1 in May-July 1979 and July-September 1983,0 otherwise; PREELECT is a dummy variable equal to 1 in February- April 1979 and April-June 1983 and 1987,0 otherwise; and D8387 is a dummy variable equal to 1 from July 1983 to June 1987, 0 otherwise. The equation suggests, first of all, that there is an inertia component to expectations about the winner of the next election. Some individuals through general lack of attention to politics, a failure during specific periods to absorb new information about the parties, or a heavy discounting of new information for partisan or other reasons fail to reassess their expectations (or do so only after some delay). The lagged dependent variable taps this inertia. In addition, one might include an inertia component in the expectations formula as used here because of the form of the measurement available to us. Our dependent variable reflects the aggregation of individual responses to a question allowing only nominal choices of the most likely winning party. Yet we would hypothesize that individuals form their expectations on a continuous scale (e.g., from 0 to 100). Thus, when asked about the most likely winner of the next election, one might considerably reassess the subjective probability that a party will defeat its opponents without crossing the 50% threshold. Such individuals, constrained by the format of the question in the Gallup surveys, will appear unchanged even though they have thoroughly reassessed a party's chances. Thus, in addition to capturing inattentiveness to political events, the lagged term will reflect hypothesized fine changes that are masked by the available measurement. The second term in the expectations equation reflects the fact that electoral expectations including lagged effects are not "time-free." They are, on the contrary, "time-bounded" by the event about which they are formed, namely, the next election (see, e.g., Lewis-Beck and Skalaban 1989, ). One aspect of this timing is that voters adopt a different perspective in forming their expectations in the months immediately after an election. To illustrate this point, consider the 1983 election. In the last Gallup poll available, fully 87% of the respondents believed that the Tories would win the election, and only 5% thought Labour would win. (WCL, our dependent variable, reaches its highest point, 82%, at that time.) In the month immediately following the election, the same percentages were, respectively, 49 and 16. After an election, voters clearly keep in mind that a new ball game has begun. To represent this discontinuity in expectations after the election, we included an interactive term, POSTELECT * WCL t _!, in our model. 18 Of greatest interest in the expectations equation are the next two variables, which together constitute an adjustment for current voting intentions. Theoretically, we are adopting here the portion of the rational expectations view that says that individuals "use current available and relevant information in forming their expectations and do not rely purely upon past experiences" (Shaw 1983, 47). Specifically, we are 373

4 Expectations in British Elections June 1994 assuming that current information about vote intentions affects perceptions about which party is likely to win the next election. 19 Of course, in a literal sense individuals cannot be reacting to the results of a survey that is only then being conducted. But people do have very recent, even simultaneous, bases for their answers. Information about candidates, parties, and politics bombards voters on a daily basis. Some of it comes directly in the form of news reports from the media; some of it is filtered through friends and acquaintances. Some of it is directly in the form of horse-race reports; 20 some of it is in the form of stories about candidates' characteristics and activities. Any single story, of course, may reach only a small minority of the population, and only a fraction of those may be persuaded to update their expectations about which party is likely to win the next election. Thus, as indicated, there is a degree of inertia in people's expectations. But along with this inertia, very recent and contemporary events and reports are indeed likely to be incorporated into current expectations. 21 In addition, individuals may reason that their current evaluations of the parties are likely to be shared by others. For example, a Labour identifier dissatisfied with his party's leaders and policies to the point of expressing a preference for the Tories may quite reasonably believe that other Labourites feel the same way (and that Conservatives, on the other hand, feel good about their party) and as a consequence feel that his party's electoral prospects are quite poor. The correspondence over time of voting intentions and expectations may thus be the result of a quite sophisticated treatment of political information. Of course, simple projection may occur as well, as voters unconsciously align their expectations with their preferences. To this degree, also, expectations will mirror the current proportions intending to vote for each party. Unfortunately, it may be impossible ever to separate these processes completely, even with individual-level data. We can show quite convincingly, however, that what we are tapping goes well beyond that kind of projection. 22 A hypothesis of simple projection would make no allowance for variations between election and interelection periods. But there are, as noted, strong theoretical reasons to believe that the process of forming expectations will have a different character in the neighborhood of the forthcoming election. First, as election day approaches, information about the parties' standings becomes a cheap and easily available commodity. Second, the incentive to use this information increases, as some voters might want to engage in tactical voting. If the degree of connection represents more than mindless projection, the electoral period should be characterized by a more pronounced disconnection between preferences and expectations as voters adjust their views about party's winning chances more quickly and more sensitively to relevant political information than in more quiet, nonelection times. That wishful thinking is harder to maintain as the election approaches can be seen in Figure 1, which show the relationship between voting intentions and expectations during the nine months prior to each of the general elections in Britain between 1966 and In every case, one finds the expected disconnection between votes and expectations, as even loyal partisans are forced to admit that they will lose their electoral bid. 24 The form, duration, arid "spread" of the party's chances of winning vary from one election to the other, but the graphs seem to show a consistent pattern, one that would not be evident if projection was the only force at work. To take this pattern into account, we have included an interactive term, PREELECT * VCL t, 25 to capture the disconnection between preferences and expectations during the electoral period. 26 This disconnection should manifest itself in a positive coefficient for the PREELECT * VCL t variable. To understand this, consider an example in which the Conservatives have the support of 50% of the electorate and Labour the support of 40% during the interelection period (VCL = 10) and in which there is a great deal of wishful thinking. Conservative and Labour supporters alike may suppose their party will win, suggesting a coefficient of approximately 1 between vote intentions and expectations of which party will win, so that net of other influences, WCL should be approximately 10 as well. Now suppose that as the election approaches, there is no change in vote intentions: it is still the case that the Conservatives have the support of 50% of the electorate and Labour the support of 40%. Those who support the Conservatives have no incentive to change their minds about which party should win the election; all 50% should still expect a Conservative victory. However, the connection between vote intentions and expectations should weaken for Labour supporters. Some of the 40%, despite their preference, will concede that the Conservatives are likely to win. This weakening of the connection between preferences and expectations means that a given vote-intention difference between the Conservatives and Labour (VCL remains 10) should translate into a larger proportion of individuals expecting the advantaged party to win (WCL becomes greater than 10); that is, the coefficient of the interaction term should be positive. 27 The two remaining variables in the expectations model have similarities to adjustment and inertia components, respectively. First, we include prospective economic thinking because the link between the state of the economy and party popularity has been demonstrated repeatedly by scholars in any number of countries. And inasmuch as these judgments are about the next 12 months, they are less subject to the problems we pointed to earlier in justifying the exclusion of unemployment and inflation variables. In addition, forecasting on the basis of future personal economic situations makes sense both in terms of retrospective and prospective thinking. If one forecasts outside the electoral period, the performance of the economy over the next 12 months is relevant 374

5 American Political Science Review Vol. 88, No. 2 FIGURE 1 Voting Intentions and Election Forecasting During Three Quarters Prior to British General Elections, March 1966 Election 100 June 1970 Election 7 8 S February 1974 Election October 1974 Election May 1979 Election June 1983 Election June 1987 Election Legend Intend to vote Conservative Intend to vote Labour Expect Conservatives to win Expect Labour to win 7 8 S

6 Expectations in British Elections June 1994 information, as it will influence the judgments of retrospective thinkers in a forthcoming election. In electoral periods as well, prospective perceptions are relevant as one may hypothesize that other voters evaluate the performance of the incumbent government in terms of the future. 28 Finally, what might be regarded as another form of inertia would manifest itself when a political party was able to establish a reputation as a winner. This reputation can develop from the subtle interplay of consecutive victories, the opposition's weakness, 30 leadership strength, 31 and past accomplishments. 32 These requirements seem to have been met during the second mandate of Margaret Thatcher as voters perceived her party as the likely winner of the next election even in light of poor showings in contemporary polls. This winners reputation component is tapped in our model by a dummy variable equal to 1 for this period. 33 As this variable intends to capture a phenomenon that is clearly specific to the expectations formation process, namely the rigidity of expectations under particular circumstances, it is not included in the vote equation. This does not mean that rigidity is without consequences on the parties' standing, as it will have an impact through the WCL variable in the vote equation. RESULTS Tables 1 and 2 show our main results. 34 With respect to vote intentions (Table 1), the first point is that the mix of economic and political variables in equation 1, along with the lagged dependent variable, appear to form an appropriate group of control variables to assess the feedback effect of expectations on voting intentions. Our strategy was to devise a vote equation that would prevent us from inflating the impact of the expectations variable, and we seem to have accomplished that goal. First, the tests reported in Appendix B for possible misspecification are reassuring on this point. Second, our results are sufficiently in accord with theoretical expectations and with other empirical results to suggest that the more standard portion of the equation yields a reasonable account of vote intentions sans expectations. As in most models of party support, for example, a lagged or autoregressive effect is clearly significant and sizable (e.g., Mishler, Hoskin, and Fitzgerald 1989). 35 That some but not necessarily all economic factors are significant is a common result (Clarke, Mishler, and Whiteley 1990, 69-80; Mishler, Hoskin, Fitzgerald 1989, 231; Norpoth 1991, ). Likewise, other studies have found only small effects of minor-party variables, whereas comparative leadership evaluations have been found to be substantial (Clarke, Mishler, and Whiteley 1990, 71; Mishler, Hoskin, and Fitzgerald 1989, 231). There are some differences between our results and those of other studies, but they are most likely attributable to different definitions of the sample period or to differences in variable definitions. TABLE 1 Voting Intentions in Britain as a Function of Economic, Leadership, System Factors, and Vote Expectations Intercept VARIABLE Vote intention (VCL) at time t - 1 a Minor party share at time t" Unemployment rate at time t - 1 Inflation rate at time t - 1 d Personal economic expectations at time t e Leadership approval at time t' Falklands War period 9 Forecast of expected winner (WCL,)* COEFFICIENT 3.69** (1.25).21** ( 06) -.06 ( 06) -.33** (.09).08 (.42).10** ( 04).27** (.03) 4.64* (2.73).10** (.02) Note: R 2 =.91, N = 138. Entries are unstandardized regression coefficients; standard errors are in parentheses. Dependent variable (VCL,) is vote intention calculated after excluding don't knows, undecideds, and "refusals"). See also n. 10. % intending to vote Cons. % intending Lab. b % minor party/(% Cons. +% Lab.). ^Monthly, adjusted seasonally. ''Monthly rate, retail price index. '% expecting better - % expecting worse. f% approval of PM % approval of opp. leader. *1 for May-June 1982, 0 otherwise. ''Instrumental variable. Instrument is the predicted value generated by regressing WCL on all exogenous variables in the system; WCL is defined in Table 2. See Appendix A for details, p <.05. "p <.01 (one-tailed test). Assuming, then, that our model is a reasonable one, we find a relatively small but clearly statistically significant effect of expectations on voting intentions. Every 10-point gap between the percentages thinking that the Conservatives rather than Labour will win the next election contributes one point to the electoral margin itself. Since 20-to-30-point gaps are not infrequent even in the interelection periods (see Figure 1) and since the margin of vote intentions is often fairly small, such an impact is far from negligible. Of course, one should not overestimate the bandwagon effect either. Even in the weeks immediately prior to the election, when the expectation gap may reach as high as 50-60%, the change in WCL from the interelection period may only account for an additional two or three percentage points. Moreover, as we shall shortly note, expectations themselves are highly dependent on vote intentions. 37 Nonetheless, it is clear that the bandwagon effect, while relatively small in 376

7 American Political Science Review Vol. 88, No. 2 TABLE 2 Forecasts of the Expected Winner of the Next General Election in Britain as a Function of Economic Expectations, Vote Intentions, and Electoral Periods VARIABLE COEFFICIENT Intercept.78 (1.16) Forecast of expected winner (WCL) at time t - 1 a Postelection * expected winner forecast (t - 1) b Vote intention (VCL,) C Preelection * vote intention (VCL,)" Personal economic expectations at time t e Conservative winning reputation'.54** (.05) -.30** (.08) 1.33** ( 18).92** (.24).21** (09) 9.39** (1.94) Note: R 2 =.91, N = 138. Entries are unstandardized regression coefficients; standard errors are in parentheses. Dependent variable (WCL t ) is forecast of expected winner. "% expecting Cons, to win - % expecting Lab. to win. ^Postelection = 1 for three months, 0 otherwise. Instrumental variable. Instrument is the predicted value generated by regressing VCL on all exogenous variables in the system; VCL is defined in Table 1. See Appendix A for details. ^Preelection = 1 for three months, 0 otherwise. '% expecting better - % expecting worse. >\ for July 1983-June 1987, 0 otherwise. p <.05. "p <.01 (one-tailed test). magnitude, may be of no small consequence in a closely contested election. Turning to the expectations equation, we noted earlier that there was less certainty about how to formulate it properly. As a consequence, all of the findings are of interest. Overall, the results (Table 2) indicate that the expectations formation process (1) exhibits both significant inertia and adjustment components, (2) reflects the voters' ability to translate expected economic conditions into political forecasts, and (3) is "time-bounded," possessing special effects both before and after the electoral period. First, note the strength of the inertia effects. To a greater extent than vote intentions, expectations about which party will win depend on prior expectations (i.e., the lagged dependent variable). An exception is in the months immediately following an election, in which the negative coefficient for the interactive term, POSTELECT * WCL t _!, means that voters recognize that the next election is a long way off and that it may therefore have little in common with what has just transpired. The consistent ninepoint bulge toward the Conservatives during the period suggests that there may be some long periods in which one party has a kind of built-in edge. Though additional time and data are required before we can explore this effect more thoroughly, we suspect that it may take on different forms in different times and places. Next note that despite these multiple sources of inertia, winning expectations are also subject to several sources of adjustment. As anticipated, there is a strong connection between current vote intentions and expectations. A 10-point gap in the proportions presently intending to vote for the two parties typically translates into a more than 13-point difference in which party is expected to win. During the months immediately preceding an election, this effect is boosted, by the coefficient on the PREELECT * VCL t interaction term, to nearly a 25-point difference. As noted, it is not possible completely to disentangle the degree to which the effect of VCL represents rational adjustment to contemporary information or blind projection. But the strength of the interaction term means that voters are able to disconnect themselves to a significant degree from their basic partisan loyalties when they perform political forecasts in the neighborhood of a forthcoming election. At least in the context of relative informational clarity of the immediate preelection period, losers as well as winners generally reach the same conclusion about the identity of the likely winner of the next election. Finally, the significant coefficient for personal economic experiences, PEXP, indicates that election expectations, like vote intentions, are tied to projections of future economic conditions. In general, optimistic economic forecasts boost the incumbent's expected chances of reelection. While the precise form of this relationship may vary from time to time, the general message is that voters are able to link economic forecasts and expectations much as they link economic forecasts with their own vote intentions. CONCLUSION Heretofore, most evidence about the interplay between election preferences and expectations has been derived from the study of American primary elections. The emphasis has been on the effects of expectations on preferences, as well as on the role of week-to-week events and of publicity in the mass media (Abramson et al. 1992; Bartels 1987, 1988; Brady and Johnston 1987; Shanks et al. 1985). This emphasis makes considerable sense in the context of primary elections, in which the contests are between individuals, not parties, the candidates are often poorly known at the start, and elections are held on a rolling basis. Though some of the same forces are obviously at work, one cannot readily generalize from these studies of American primaries to the nature of expectations in general elections. Therefore, we developed a model to fit the latter circumstances. In doing so, we placed greater emphasis on the effects of preferences on expectations and on the nature of the expectations-formation process in less intense, interelection times as well as during election periods. 377

8 Expectations in British Elections June 1994 Our findings show, first of all, that the expectations-formation process takes on a different character during interelection times. Similarities exist between the electoral period and more quiet times (inertia exists throughout, and the economic forecast and winning reputation variables remain unchanged) but there is a major disconnection between voter preferences and expectations in the neighborhood of an election. Particularly revealing is the preelection adjustment (PREELECT * VCL t ). It may be tempting to interpret the quick adjustment component (VCL t ) simply as a manifestation of what Bartels called "the resilience of projection... even in situations where objective cues about the shape of political reality are quite powerful" (1987,811). Indeed, the fact that only the intensity of the electoral campaign introduces a strong reality check lends credence to the idea that most of the time, most of the people think that future outcomes will move in the direction of their own preferences. But the larger, combined effect of aggregate preferences and simple projection (and whatever other psychological mechanisms link current preferences and expectations) in periods close to the election tells us that reality rapidly takes over when the election is actually in sight. 39 Our findings also reveal the contemporaneous nature, even during interelection periods, of the effect of vote intentions on expectations and vice versa. Before they were even completed, much less publicized by the media, the outcomes of polls of vote intentions were already internalized by respondents and reflected in the degree to which they thought that one party held the advantage going into the next general election. Likewise, the anticipated election outcomes had already influenced, though to a smaller degree, those same vote intentions. While the contemporaneous impact of vote intentions on expectations is consistent with the notion of a simple projection effect, it also suggests how readily available is information about party popularity and related matters. Despite the little attention that ordinary voters pay to political matters, the aggregate movements studied here appear to reflect an ongoing awareness of and reaction to the political world. More generally, the results are also consistent with the view that people accommodate political information quickly and efficiently. In making much the same point, Lewis-Beck and Skalaban suggested that about all that voters needed to perform competent electoral forecasts was "regular membership of the [American] polity" (1989, 153). Our own work goes one step further by suggesting that voters use quite simple forecasting rules, weighting, when it is relevant, the burden of the past (the winning reputation component) with current facts and evaluations about who they support, what they expect their future financial situation to be, and when the next election will occur. With respect to vote intentions, our finding of a bandwagon effect, even though small, adds support to the contention that some voters respond to the implicit bonus of being on the winning side even in high-salience general elections. Bandwagon effects have been found primarily in U.S. presidential primaries (Abramson et al. 1992, 64-65; Bartels 1988; Skalaban 1988,136-50), but as noted, those elections represent a very different setting from most electoral situations. In British studies, evidence consistent with bandwagon effects has been found, but the absence of panel data made it impossible to distinguish with certainty between bandwagon and projection effects (McAllister and Studlar 1991,736). And in a Canadian study, there was almost no evidence of bandwagon effects (Johnston et al. 1992). It may well be that their small size not only makes bandwagon effects difficult to detect but that they are sometimes swamped by other election factors. At any rate, the existence of a bandwagon effect in our study suggests that in the future, properly specified popularity functions should include an expectations term whenever possible. Finally, the results of the overall model are refreshing from a normative point of view. They indicate that many among the "losers" are able to make reasonable forecasts without being in turn heavily influenced by this disappointing outcome. That so many losers recognize themselves as such and nevertheless stick to their preferences is refreshing in that it strongly suggests that under normal circumstances voters are expressing preferences and not simply following the crowd. APPENDIX A: IDENTIFICATION AND ESTIMATION OF THE EXPECTATIONS AND VOTE MODELS This appendix gives details of the identification and estimation of our equations. Identification Identification of our model is complicated by the fact that we have a system of equations that is nonlinear in the variables. In particular, the expectations equation contains a current endogenous variable on the right-hand side that enters in both an additive and an interactive fashion. It is well known that the usual necessary-and-sufficient rank condition for identification of linear structural equations needs to be amended in this context (Achen 1986,141-44; Greenberg and Webster 1983, ; Kelejian and Oates 1989, ). Fortunately the required rank condition is a simple extension of the linear case. In a general linear simultaneous-equations model with G endogenous variables, the rank condition requires that for each equation in the system rank (A <E>) = G - 1 where A = [T B] is a matrix comprised of the coefficients on the endogenous and exogenous variables in the system and 4> is a conformable matrix representing the exclusion restrictions present in the system. In the linear-in-the-variables case, we require that the endogenous variables enter the A matrix 378

9 American Political Science Review Vol. 88, No. 2 (through F) additively. This rules out models like ours where the A matrix contains an interactive term due to the specification of the expectation equation. However, a simple extension of the rule rank (A <f>) = G - 1 will cover our nonlinear-in-the-variables case. (For a general proof see Fisher 1966, chap. 5). Provided we include the nonlinear function of the endogenous variable as an additional column of A-F, we can continue to use the rule rank (A 4>) = G - 1 to check the identification status of the model. In more practical terms, provided we treat the interactive term in the expectations equation as an additional endogenous variable holding constant the number of rows of A = [F B], it is simple to check the identification status of our expectations and vote equations. The reader will find that both our equations are overidentified. Estimation The estimation of our model is a relatively straightforward application of two-stage least squares an instrumental-variables method in which the "instrument" is simply the predicted value of the endogenous variable(s) from a reduced-form regression equation. As with the identification procedure, estimation is slightly more complicated when the model is nonlinear-in-the-variables. In the linear simultaneous equations model the instrument is formed by regressing each endogenous variable in the model on all exogenous variables in the system. In the nonlinear case the instrument is often formed by regressing each endogenous variable on all the exogenous variables plus a set of polynomials and interactive terms in the exogenous regressors (see, e.g., Achen 1986, ). We computed the instrument by the standard method, eschewing the more complex nonlinear-in-the-variables reduced form (although we tried this specification and it made little difference in the results). A justification for our choice of procedure can be found in Kelejian and Oates 1989, Provided the nonlinearity is not too severe (and our reduced-form estimates indicate that this is the case), the linear approximation in the instrument equation will not cause problems in formulating the instrument. As noted, the actual estimation procedure is twostage least squares, which is simply a particular form of instrumental variables. The asymptotic theory that underlies estimation and inference in linear and nonlinear expectations models can be found in Pagan 1984, Our model is a variant of Pagan's (1984) model 2 with some additional attention being paid to the nonlinearity. Standard asymptotic theory on the convergence of the normalized moment matrix of regressors and the cross-moment matrix of regressors and disturbances leads to consistency results in the location (slope) parameters. As is standard, the residuals are calculated from the original variables leading to consistency for the scale (variance/standard error) parameter. Proofs are straightforward and can be found in Pagan 1984, (the linear model) and Pagan 1986, (the nonlinear model). The diagnostics, which indicate that our specification is sound, can be found in Appendix B. APPENDIX B: DIAGNOSTIC TESTING PROCEDURES To check for general adequacy of our model specification we used the diagnostic testing procedures advocated by Spanos (1986) and applied in political science by Granato (1991). The PC-GIVE package provides a series of regression diagnostics designed to alert one to general model misspecification. The tests provided include RESET tests for functional misspecification, a Lagrange multiplier test for general patterns of serially correlated errors, a test for ARCH (Autoregressive Conditional Heteroscedasticity) patterns in the residuals, a test for general heteroscedasticity in the residuals (the "White" test), and a test for residual normality. Because we are operating in a simultaneous-equations framework, we use the residuals based on the original variables, not their predicted values. The relevant statistics based on these residuals converge to the chi-squared distribution, as do their single equation counterparts (Spanos 1986, chaps ). The substantive aspects of our regression diagnostics follow. First, our model appears to be well specified: the RESET test for both vote and expectation equations indicates that there is no obvious misspecification in terms of functional form or omitted variables. The relevant statistics fall well below conventional levels of significance (x 2 5 = 3.30 for the expectations equation, and Xi' = 1-53 for the vote equation). Second, serially correlated errors do not plague our analysis: given that our model includes a lagged dependent variable, the standard Durbin-Watson statistic is not very meaningful. A more appropriate and general test is the Lagrange multiplier test. Using PC-GIVE, we performed a joint test on lags 1-7 of the autocorrelation coefficient. The resulting statistics fall well below conventional levels of significance (x 2 7 = 1.49 for the expectations equation, and # 2 7 =.22 in the vote equation). Additional evidence confirming the absence of serial correlation can be seen by looking at the magnitude of the autocorrelation coefficients at lags 1-7. No estimated coefficient is larger than.20, and most of them are in the -.05 to.05 range. There does appear to be extremely mild heteroscedasticity for the expectations equation (xi 2 3 = 22.96, significant at the.05 but not at the.01 level) and no visible traces of it in the vote equation (x 2 16 = 14.56). The evidence on lack of heteroscedasticity, while not quite as good as that on autocorrelation, seems to be reasonable. Finally, both the expectations and the vote equations have residuals that pass a x 2 test for normality 2 (x = for expectations and Xz = 5-61 for vote). In summary, model diagnostics indicate that our two-equation model is theoretically well specified, 379

10 Expectations in British Elections June 1994 \ does not have serious problems in the stochastic component, and possesses residuals that are normally distributed. Notes We would like to thank Kirke Lawton for excellent technical assistance. We also wish to thank Neal Beck, Helmut Norpoth, and William Keech for their comments on a draft of this article. 1. The voting intention question is, "If there were a General Election tomorrow, how would you vote?" The expectations question is, "Regardless of how you are going to vote yourself, which party do you think is most likely to win the next general election?" 2. The expectations question was asked irregularly (mostly in months just preceding an election) from 1945 to Since 1975, it has been asked every month except seven months after the 1979 elections and five months after the 1987 elections. We limit our attention to the period starting in January 1976 and ending in June 1987, for two basic reasons: (1) including the period after the 1987 election would have been demanding in econometric terms as it would have created a second string of missing data; and (2) since the popularity function of the British parties has been extensively studied for different subperiods up to the end of the second mandate of Margaret Thatcher, it will be easier to specify a satisfactory vote equation, a preliminary condition in any study about the links between votes and expectations. On the popularity function, see Clarke, Mishler, and Whiteley 1990; Mishler, Hoskin, and Fitzgerald 1989; Norpoth 1991; Sanders, Ward, and Marsh The original expressions of both preferences and expectations are categorical rather than thermometers measuring intensity of feelings and estimated probabilities. In addition, we have available only aggregate rather than individual-level preferences and expectations (though aggregate data may well be better than individual data, as noted in Kramer 1983 and Lewis-Beck and Skalaban 1989, 153). 4. We shall sometimes use the terms votes and votes equation rather than the more precise but longer terms voting intentions or popularity function. Similarly, we shall use expectations about future elections and election forecasts interchangeably. 5. We formulate the model in terms of general elections. So far, the most intense scrutiny of expectations formation has been in U.S. primary elections. We discuss the differences between those models and our own in the conclusion. 6. VCL, WCL, LEADERS, and MPSHARE are taken from Butler and Butler 1986 updated to 1987 from the Gallup Poll Monthly, UNEMP and INF are taken from the Organization for Economic Cooperation and Development's Main Economic Indicators: Historical Statistics, PEXP is taken from the Gallup Poll Monthly. As indicated in n. 2, WCL was unavailable for seven months during the period under study. We estimated our models both with and without fitted values (computed from the model's parameters) for the missing data and obtained nearly identical results in terms of the parameters and the goodness-of-fit measures. Results with fitted values are reported here. 7. For the purposes of estimation, the endogeneity of WCL, and VCL t poses a problem when they appear as explanatory variables in the vote and expectation equations, respectively. For this reason, an instrumental variable in the form of a predicted value is created when either variable appears on the right-hand side of our equations. See Appendix A for details. 8. Inclusion of both objective conditions and personal expectations has been common in the popularity function literature in Great Britain ("A Computer's Election Advice," Economist, 17 August 1991, 53; Clarke, Mishler, and Whiteley 1990; Sanders, Ward, and Marsh 1991). Some work on U.S. presidential popularity argues in favor of including only > retrospective (Beck 1989) or only prospective (MacKuen, Erik- son, and Stimson 1992) economic variables. Our interest here > is not in testing these alternative views but to take reasonable ' precautions against overestimating the effects of expectations. 9. How do you think the financial situation of your household will change over the next 12 months? {Get a lot better. Get a little better. Stay the same, Get a little worse, Get a lot worse). Following Clarke, Mishler, and Whiteley 1990 and Sanders, Ward and Marsh 1991, we operarionalized the variable as the percentage difference between the first two categories (a lot better or a little better) and the fourth and the fifth categories (a little worse or a lot worse). 10. Note also that both variables (as well as the subjective perception and the LEADERS variables) were designed in accordance with a reward-punishment model. To do so, "we reverse the sign of unemployment and inflation during Conservative governments thereby predicting that the Conservative... [lead] will increase both: a) as inflation and unemployment rise during Labour governments and b) as inflation and unemployment decline during Conservative governments" (Mishler, Hoskin, and Fitzgerald 1989, 224). Note finally that the rival hypothesis, the issue priority model, was also examined. (We also checked the possibility that the Tories may have benefitted from an "incumbency effect" by adding a dummy variable taking the value of 1 for the two Conservative mandates. It was not significant.) A useful summary of the differences between the issue priority model and the standard reward-punishment model is Clarke, Stewart, and Zuk 1986, See also Kiewiet 1983; Kiewiet and Rivers 1984; Mishler, Hoskin, and Fitzgerald 1989, ; Peffly Retrospective perceptions about both personal and general economic conditions, as well as a variable measuring prospective perceptions about general economic conditions, were also examined. 12. Lafay, in a French study, follows a rather similar practice by introducing the popularity index of the prime minister in an equation for the popularity function of the president (1991, 135). Micro analyses also show that evaluations of party leader performance in Britain have independent effects on party support (see Clarke, Stewart, and Zuk 1986; Graetz and McAllister 1987). For evidence that leader images in Britain affect not only short-term intentions such as voting but also long-term intentions such as party identification, see Clarke and Stewart There has been a debate about the magnitude and form of the impact of the Falklands War on parties' standings. Consequently, we tried various specifications for this variable and selected the best one using standard criteria. We think it likely that our simple dummy variable, together with the LEADERS variable, adequately takes into account the immediate Falklands War effect and its lasting impact through the sustained popularity of Mrs. Thatcher. On the debate over the war's effects, see Clarke, Stewart, and Zuk 1986; Clarke, Mishler, and Whiteley 1990; Norpoth 1987a and b; Sanders, Ward, and Marsh Specifically, one would write down a structural model of the economy and use all the variables in the system to help predict the expected value of the endogenous variable of interest. Technically, this would impose a series of nonlinear cross-equation restrictions between the "reduced form" equation and the "expectations" equation. This is not the strategy we pursue here, for reasons outlined in the text. 15. This would be especially true if the data were "difference-stationary" rather than "trend-stationary." However, in our model only the inflation variables might reasonably be viewed as a difference-stationary process (but see Ostrom and Smith 1993). 16. First, the component satisfaction with the prime minister is basically retrospective. Second, the question about the opposition leader is also retrospective in that it does not ask the respondent to think about the opposition leader as a potential prime minister. Third, the two components do not tie the leaders together in an explicit inquiry about their comparative 380

11 American Political Science Review Vol. 88, No. 2 potential as future prime minister. The question, "Who would make the best prime minister?," followed by a list of potential leaders, would perhaps be a better choice in an expectations equation, but it has not been asked on a regular basis. An example will illustrate these various points. In December 1985, satisfaction with Prime Minister Thatcher was 39%, with Kinnock at 47%; yet when asked who would make the best prime minister, Thatcher received the support of 33%, Kinnock of 24%. 17. Our expectations model is consistent with the argument outlined in Beck Beck argues that one ought to think of economic forecasts as conditional expectations where the conditioning variables include competence, administration, and other political variables. He goes on to criticize the strong rational-expectations approach in the popularity-function literature (e.g., Chappell and Keech 1985; Chrystal and Peel 1985) and presents empirical evidence indicating that such a model is suspect. (See also the criticisms of the rational-expectations approach in Hudson 1982, chap. 3.) We believe that our expectations model, which is more than an adaptive expectation but less than the economists' rational expectations, is an appropriate forecasting model. 18. This approach is similar to that used in studying presidential popularity functions, in which some account needs to be taken of changes in administrations. See, e.g., Beck 1991, 91, n. 11). 19. The high correlation between current intentions and expectations (.80 for the period under study) supports the view that individuals update their views about the likely winner of the next election very quickly, even continuously. 20. Evidence about the availability of poll results can be found in Butler and Kavanagh Evidence about public exposure to polls for the elections under study here can be found in McAllister and Studlar 1991, Indeed, to exclude VCL t from the equation would require, instead, that we include variables representing sources of all the new information used to form the voter's current vote intention. It is likely, of course, that some voters are sufficiently insulated from contemporary information that they seldom update their understanding of voting intentions. Our argument is that on balance, the most immediate measure of vote intentions (i.e., at time t) is a superior predictor of current expectations. We tested this reasoning, however, by trying an alternative specification in which VCL,_i replaced VCL t. The resulting coefficient was not significant despite the high correlation (.78) between the two vote measures. 22. In fact, if simple projection from one's own vote intentions were the only process involved, then only individuallevel VCL would be appropriate in the expectations equation. However, our interpretation is much broader. As we noted, individuals probably feel that their current evaluations are shared by others, but surely they do not think that everyone else views the parties identically. Thus, it seems fully justifiable to include aggregate-level VCL in the expectations equation. The interaction of VCL with PREELECT is also a strong indication that aggregate VCL belongs in the expectations model. 23. We expand our time frame here because data for nine months prior to each election are available for the longer period. 24. Note that voters put great reliance on polls to form their expectations even if they were misleading, as in 1970 and Do the additive variable PREELECT and the variable POSTELECT (introduced earlier) need to be included as well for proper specification? Not only is it unnecessary always to include each of the components of an interactive term (see, e.g., Lewis-Beck 1980, 54-56), but it is relatively meaningless in this case, as the correlation between PREELECT and PREELECT VCL is.93, and that between POSTELECT and POSTELECT * WCL is.63. As to the choice of which variables to include (additive or interactive), theory clearly suggests the interactive form. If two parties are perceived as having exactly the same support, there is no reason to believe that the type of "blips" illustrated in Figure 1 will occur (as a dummy variable would suggest). Moreover, a dummy variable would imply that these blips will have the same magnitude whatever the vote difference between the main contenders. An interaction term is more flexible, as it allows the sudden increase of "nonprojecting" voters to be linked to the closeness of the election. 26. Throughout, the election period is operationalized as the two months preceding an election as well as the electoral month itself (with a slight adjustment in 1979, when the election was on 3 May). Though electoral campaigns as such are shorter than that in Great Britain, rumors and polling activities justify this definition. The graphs show that the election effect does not occur smoothly over the interelection period but rather abruptly in the neighborhood of a forthcoming election presumably under pressure of the unmistakable realities of the electoral campaign, the enhanced attentiveness of the public, and the decreased likelihood that time will permit a trailing party to rebound significantly. The uniqueness of the electoral context also explains the abruptness of the postelection adjustments we have discussed. 27. That a comparable interaction term would be negative in a model of individual vote decisions is entirely consistent with our argument, as it is precisely the weakening of projection effects (during election periods) among those who are likely losers that will generate the widening of the gap between preferences and expectations illustrated in Figure 1 and in the example in the text. 28. We checked for a possible interaction between time (operationalized both as one year before the election or not and number of months until the next election) and PEXP and found it not significant; perhaps the effect of this variable is about the same in election and nonelection periods precisely because it can be relevant for both retrospective and prospective reasons. 29. The two consecutive victories by Margaret Thatcher followed a period during which the incumbent party had been defeated four times out of six (1964, 1970, 1974, and 1979). Moreover, her 1983 victory was the most impressive one in the postwar period both in terms of seats and the popularity gap with Labour. Conversely, Labour's performance in this election was the poorest one in seats and popular support for the same period. 30. Labour was perceived during this period as much more divided than the Tories. In fact, in all the surveys during Thatcher's second mandate, the percentage of respondents perceiving Labour as a divided party outnumbered by extremely large margins the percentage of those thinking of it as united (-55 on average). The Tories were basically perceived as a united party for the same period, with exceptional and short periods of negative balance, (see Gallup Poll Monthly). 31. The high personal popularity of Thatcher, as well as the Tory's advantage in questions about the best leaders, speak to this point (see Gallup Poll Monthly). 32. The victorious Falklands War, the decrease of the inflation rate, and the economic recovery are possible candidates here. 33. That we found out-of-sample confirmatory evidence of the rigidity of expectations during Thatcher's third mandate buttresses our confidence that the type of inertia measured by our winning reputation variable is a genuine characteristic of expectations. Moreover, a preliminary inspection of the data after the Conservatives' fourth consecutive victory in 1992 apparently shows less rigidity, which is also consistent with our hypothesis that a string of victories is a necessary though not a sufficient condition to produce rigidity in expectations. 34. R 2 is estimated as 1-Regression Sum of Squares/Total Sum of Squares. 35. Should we formally test for integration of order one (1(1)) or a "stochastic nonstationarity" in the data series? We feel that such integrated behavior is highly unlikely, given the departure of our lagged coefficients from unity and the fact that our data are rates or ratio-type variables. For a discussion of this point and some general problems with the unit root methodology involved, see Cochrane 1991,

12 Expectations in British Elections June The reported effect of the Falklands variable may seem small compared to other work. However, it is important not to lose sight of two facts. First, the results will be different depending on the definition of the dependent variable. The war had a different impact on Thatcher's approval rate (immediate and lasting) from what it did on the Tories' popularity (immediate but rapidly declining). Second, the results will be greater if the approval rate of Thatcher is excluded from the regressors rather than included, as here. 37. In addition, we tested for the possibility that bandwagon effects were greater in the period immediately prior to elections (by including an interactive term, PREELECT * WCL). The coefficient was not significant. Following Norpoth, we also checked the possibility that electoral campaigns drive disenchanted partisans back to the governing party (1991, ); we did so by adding separate dummies for the electoral periods under study. None of these variables was significant. 38. For example, in a Canadian study, a simple counting variable denoting the number of consecutive polls in which a given party had led in preelection polls was important. We tried such a variable here, but it was not significant (perhaps because of the much greater time between polls, which meant that winning streaks were probably remembered less well, and, in any event, were not so long lived (in terms of numbers of polls) as those in Canada. We also tried a variable measuring the proportion of times during the last six months in which the Conservatives were leading in the polls, but it was not significant (see Johnston et al. 1992, chap. 7). 39. Sometimes, reality becomes overwhelming, and "almost everyone" correctly forecasts the election result, as in Britain in 1983 (see Figure 1) or in the U.S. elections of 1964 and 1972 (Lewis-Beck and Skalaban 1989, 148). Of course, even in the case of a landslide, there will be some "true believers." Moreover, a more adequate measure of expectations (0-100 scale) would probably capture projection effects not visible with the measure we have available; for example, in the U.S. presidential election of 1964, a staunch Republican might have given Goldwater a 25% chance of winning, while most others would have given him close to a 0% chance. 40. These authors also point out that British work on this topic has produced conflicting results; see McAllister and Studlar 1991, 724. That work has often been based on simple comparisons between election results and public opinion polls conducted during the campaign. References Abramson, Paul R., John H. Aldrich, Phil Paolino, and David W. Rohde " 'Sophisticated Voting' in the 1988 Presidential Primaries." American Political Science Review 86: Achen, Christopher H The Statistical Analysis of Quasi- Experiments. Berkeley: University of California Press. Alt, James E The Politics of Economic Decline. Cambridge: Cambridge University Press. Bartels, Larry "Candidate Choice and the Dynamics of the Presidential Nominating Process." American Journal of Political Science 31:1-30. Bartels, Larry Presidential Primaries. Princeton: Princeton University Press. Beck, Nathaniel "Presidents, the Economy, and Elections: A Principal-Agent Perspective." In The Presidency in American Politics, ed. Paul Brace, Christine Harrington, and Gary King. New York: New York University Press. Beck, Nathaniel "The Economy and Presidential Approval: An Information Theoretic Perspective." In Economics and Politics, ed. Helmut Norpoth, Michael S. Lewis-Beck, and Jean-Dominique Lafay. Ann Arbor: University of Michigan Press. Begg, David K. H The Rational Expectations Revolution in Macroeconomics: Theories and Evidence. Oxford: Philip Allan. Brady, Henry E., and Richard Johnston "What's the Primary Message: Horse Race or Issue Journalism?" In Media and Momentum, ed. Gary Orren and Nelson Polsby. Chatham, NJ: Chatham House. Butler, David, and Gareth Butler British Political Facts, Basingtoke: MacMillan. Butler, David, and Dennis Kavanagh The British General Election of London: MacMillan. Cagan, Philip "The Demand for Money in a Hyperinflationary Economy." In Studies of the Quantity Theory of Money, ed. Milton Friedman. Chicago: University of Chicago Press, Chappell, Henry, and William Keech "A New View of Political Accountability for Economic Performance." American Political Science Review 79: Chrystal, Alec, and David Peel "What Can Economists Learn from Political Scientists and Vice Versa?" American Economic Review 76: Clarke, Harold D., William Mishler, and Paul Whiteley "Recapturing the Falklands: Models of Conservative Popularity." British Journal of Political Science 20: Clarke, Harold D., and Marianne C. Stewart "Dealign- \ ment of Degree: Partisan Change in Britain, " j Journal of Politics 46: I Clarke, Harold D., Marianne C. Stewart, and Gary Zuk i "Politics, Economics, and Party Popularity in Britain, " Electoral Studies 5: Cochrane, John "Comment." In National Bureau of Economic Research Macroeconomics Annual 1991, ed. Oliver Blanchard and Stanley Fischer. Cambridge: Massachusetts Institute of Technology Press. Dunleavey, Patrick, and Christopher T. Husbands British Democracy at the Crossroads. Boston: Allen & Unwin. Fisher, Franklin M The Identification Problem in Econometrics. New York: McGraw-Hill. Frey, Bruno S., and Friedrich Schneider "A Politico- Economic Model of the United Kingdom." Economic Journal 88: Goodhart, C. A. E., and R. J. Bhansali "Political Economy." Political Studies 18: Graetz, Bernard, and Ian McAllister "Party Leaders and Election Outcomes in Britain, " Comparative Political Studies 19: Granato, Jim "An Agenda for Econometric Model Building." Political Analysis 3: Greenberg, Edward, and Charles E. Webster, Jr Advanced Econometrics. New York: Wiley. Hibbs, Douglas A., Jr "On the Demand for Economic Outcomes: Macroeconomic Performance and Mass Political Support in the United States, Great Britain, and Germany." Journal of Politics 43: Hudson, John Inflation: A Theoretical Survey and Synthesis. London: George Allen & Unwin. Hudson, John "The Relationship between Government Popularity and Approval for the Government's Record in the United Kingdom." British Journal of Political Science 15: Johnston, Richard, Andre Blais, Henry E. Brady, and Jean Crete Letting the People Decide. Montreal: McGill- Queen's University Press. Kelejian, Harry H., and Wallace E. Oates Introduction to Econometrics. 3d ed. New York: Harper & Row. Kiewiet, D. Roderick Macroeconomics and Micropolitks. Chicago: University of Chicago Press. Kiewiet, D. Roderick, and Douglas Rivers "A Retrospective on Retrospective Voting." Political Behavior 6: Kramer, Gerald H "The Ecological Fallacy Revisited: Aggregate- Versus Individual-level Findings on Economics and Elections and Sociotropic Voting." American Political Science Review 77: Lafay, Jean-Dominique "Political Dyarchy and Popularity Functions: Lessons from the 1986 French Experience." In Economics and Politics, ed. Helmut Norpoth, Michael S. Lewis-Beck, and Jean-Dominique Lafay. Ann Arbor: University of Michigan Press. j 382