Macrointerest

Abstract An interested and engaged electorate is widely believed to be an indicator of democratic health. As such, the aggregate level of political interest of an electorate – macrointerest – is an essential commodity in a democracy, and understanding the forces that change macrointerest is important for diagnosing the health of a democracy. Because being interested in politics requires time and effort, the article theorizes that the electorate's level of political interest will be highest when the electorate believes the government cannot be trusted or is performing poorly. To test hypotheses derived from a proposed theory against rival explanations, the study develops a measure of macrointerest using a quarterly time series of aggregated survey items (1973–2014) of political interest. The authors find support for the theory that the electorate responds as reasonable agents when determining how closely to monitor elected officials: interest is positively related to decreases in trust in government.

Interest in politics has been foundational in the study of political behavior. The American Voter, The People's Choice, The Responsible Electorate, The Semi-Sovereign People, and other important works recognize and elaborate the powerful role that interest in politics plays in American political behavior (for example, Campbell et al. 1960;Key 1966;Lazarsfeld, Berelson and Gaudet 1948;Schattschneider 1960). Citizens who are interested in politics are more knowledgeable (Delli Carpini and Keeter 1996) and more likely to engage in a multitude of political activities than their less interested counterparts (Verba, Schlozman and Brady 1995). Models predicting just about any form of political participation that exclude interest in politics are simply misspecified.
Political interest is more than just a prominent predictor of political participation. Dating back to Almond and Verba (1963), the political interest 1 of the electorate has been characterized as an important aspect of a country's civic and political culture; its presence or absence has since been thought of as an indicator of a democratically stronger electorate. In the comparative politics literature, aggregate 2 levels of political interest are viewed as a measure of the health of a democracy. Electorates that pay more attention to the goings-on in government are believed to signal stronger democracies, and declines in political interest are used as evidence of democratic deconsolidation (Foa and Mounk 2016). The strength of a citizenry's ability and/or motivation to engage in democratic accountability is often thought to depend on the levels of interest and attention that citizens pay to government (Solt 2008).
A more interested electorate exhibits the kinds of basic civic behavior and responsibility that are believed to promote stronger political and democratic institutions (Prior 2018). As such, understanding the levels and dynamics of the electorate's political interest should provide evidence of the health of a country's democracy. However, we simply have no idea what features of American politics lead the American public to be more or less interested in politics over time.
The lack of empirical focus on the levels and dynamics of macrointerest is not because political interest is absent from theories of macropolitics in America. Although most research on political interest focuses on individual behavior, the aggregate level of political interest also plays an important, albeit implied, role in macrolevel theories of American politics. Much of the macropolitics literature focuses specifically on the mass public, concluding that, when the public chooses to register its views, public opinion is a more significant factor in policy making. For example, Erikson, MacKuen and Stimson (2002) find that there is a dynamic relationship between the public and the political system. They argue that the public's preferences constrain the political system, which in turn affects the relative preferences of the public. Page and Shapiro's The Rational Public (1992) is predicated on the notion that the public is able to follow politics closely enough to acquire information, but they note that changes in the public's overall ability to acquire information about politics are likely to moderate the ability of the public to constrain the government. One precursor to the public's ability to assert its issue and candidate preferences is the degree to which the public is paying attention to and monitoring the goings-on in government and the policy environment, which we assert is captured by measures of political interest.
Along with political interest, trust in government is also viewed as an indicator of the health of a democracy. In general, democracies are associated with higher levels of citizen trust than nondemocracies (for example, Jamal and Nooruddin 2010; see Letki 2018 for a review). And trust is positively related to support for both democracy (Bratton and Mattes 2001;Kunioka and Woller 1999) and progressive policies (Hetherington 2005). The more people trust government, the more willing they are to support policies that expand its scope. Conversely, when trust is low, the public prefers policies that limit government's reach. 'When government programs require people to make sacrifices, they need to trust that the result will be a better future for everyone' (Hetherington 2005, 4). Similarly, Warren (2018) argues that 'Well-functioning democracies are able to provide common goods for themselves, which requires high levels of public trust in those parts of government with missions to provide these kinds of goods ' (89, emphasis in original).
Trust in government is also positively associated with political participation (Dalton 2004;Smets and Van Ham 2013; see Hooghe 2018 for a review), and trust in justice institutions is an important predictor of co-operation and compliance with legal authorities (Bradford, Jackson and Hough 2018). With regard to public opinion, political trust acts as a heuristic that enables people to decide whether or not to support new government policies or initiatives, as well as the government's projections about the future consequences of policy proposals (Hetherington and Rudolph 2015).
Given the positive effects of trust in government on indicators of democratic health, the decline in trust over the past few decades has caused a great deal of academic handwringing over the state of American democracy (see Citrin and Stoker 2018 for a comprehensive review). The fall in the electorate's trust in the federal government is a ubiquitous symbol of all that is wrong with American politics, and is believed to be one of the most fundamentally troubling trends of the last half-century.
To date, no systematic analysis of the relationship between these two fundamental constructs related to the health of democracypolitical interest and political trusthas been conducted. The goals of this manuscript are therefore threefold. First, we develop a theory that connects political interest and trust in government to delineate the conditions under which political interest should ebb and flow. We argue that, at the macro level, decreases in trust in government should cause increases in political interest. Secondly, we develop a measure of aggregate political interest (what we term 'macrointerest') using an extensive aggregated quarterly time series  of political interest public opinion questions and test our theoretically derived hypotheses about the antecedents of macrointerest against those of alternative theories. Finally, we discuss the implications of our findings for understanding the role of aggregate political interest in a democracy.

From Micro to Macro
Despite its prominence in the field, there is a surprising lack of attention to the antecedents of political interest. At the individual level, much of the seminal work cited above does uncover some of the social and political differences between the interested and the uninterested. However, other than noting that educated citizens or people whose parents were politically engaged are more likely to be interested, the field has paid scant attention to the sources of political interest at the cross-sectional, individual level, or to developing a theory of the antecedents of interest. Most recently, Prior (2018) demonstrates with a wealth of panel data that political interest is largely stable at the individual level, and therefore argues that the antecedents of microinterest occur early in life. Prior (2018) also argues that because microinterest is quite stable, macrointerest should be as well. 3 We are skeptical of this claim. Consistent with the broader macropolitics literature, we argue that there is nothing inconsistent with individual-level stability and aggregate-level volatility substantively or theoretically. For example, as research on the relative stability of party identification has demonstrated, individual-level stability does not guarantee the same at the aggregate level (Erikson, MacKuen and Stimson 1998;MacKuen, Erikson and Stimson 1989). Reconciling the two findings is relatively easy: at the individual level, if 95 per cent of respondents report the same partisanship over time, most scholars would call that stable. If, however, the remaining 5 per cent of respondents change in a systematic way, and all move in the same direction due to some political event or figure, the result would be quite a substantive swing in the country's aggregate partisan leanings. All that is required for there to be substantive and potentially consequential movement in the aggregate (that is, systematic variance that we can attempt to explain with our theory) is for even a small percentage of people who do change to do so in a uniform direction. Therefore, contrary to Prior (2018), we argue that whether or not macrointerest is as stable as microinterest is an open empirical question (and one that is beyond the scope of this manuscript). To elaborate, our goal is not to characterize the amount of variability that exists in macrointerest; nor is it to determine whether macrointerest is more or less stable than microinterest. Rather, our goal is to test our theory that political interest in the aggregate, as an indicator of the extent to which people are motivated to monitor the goings-on of government, is inversely related to aggregate trust in government. In order to test our theory, we must develop a measure of macrointerest, and provide some evidence that there is at least some variability in macrointerest that we can try to explain using our theory.
Unlike other concepts like trust in government (Keele 2005), macropartisanship (MacKuen, Erikson and Stimson 1989), ideology (Box-Steffensmeier, Knight and Sigelman 1998), presidential approval (Mueller 1970) or public policy mood (Stimson 1998), there has been no attempt to develop an aggregate-level measure of the dynamics of the electorate's changing interest in politics. This lacuna is striking, given the role that interest plays in theories about the health of democracies. An aggregate measure of political interest would enable scholars to not only test 3 The primary goals of Prior's (2018) work are to explain the individual-level causes of political interest, and to determine whether, and to what extent, individuals' interest in politics changes during the course of their lifetimes (he amasses an impressive amount of data to make a convincing case that political interest at the individual level develops in early adulthood and is quite stable over time). Prior (2018) also presents data about population-level variability in interest by charting individual cross-sectional time-series surveys (such as the American National Election Studies (ANES)). However, given that his primary focus is the individual level, his aggregate charts are purely descriptive; he visualizes the various series and concludes that because they do not move up and down all that much, aggregate interest is also relatively stable.
theoretically derived hypotheses about the causes of ebbs and flows in the electorate's political interest; it would also let them directly test the theoretical link between political interest and democratic health.
Macrointerest, our term for the American public's aggregate level of interest in politics, is important not only for its use as a metric of the health of democracy, or due to its potential role in moderating the responsiveness of the public and the government. In this article, we argue that macrointerest is also a negative consequence of one of the other leading indicators of democratic healthtrust in government (as well as, potentially, other measures of government performance such as presidential approval).

The Role of Macrointerest
Our argument about the causes of increases in the electorate's political interest relies on the theoretical perspective that views the workings of American democracy as analogous to a principalagent problem (Ferejohn 1986). In this case, the elected officials are the agents who are wont to act in their own self-interest and the mass public is the principal. Voters have preferences regarding the behavior of their agents, as well as the content of the policies they propose and implement, but most segments of the polity have limited information about the political activities of its elected agents. Additionally, the electorate often has few incentives to monitor elected officials. But when the principal is paying attention, the agent is more constrained in its pursuit of power and resources (for example, Mayhew 1974). As Anderson (2007, 287) explains, "As an empirical matter, accountability happens when voters can and want to discern whether governments are acting in their interest." As such, the questions of whether, and the conditions under which, the polity decides to pay attention to and monitor the goings-on in government are worthy of scholarly attention.
As with any principal-agent relationship, the principal needs to decide how much effort to exert to make sure the agent is performing acceptably. The degree of monitoring is likely to be a function of the signals the electorate receives about the performance of elected officials and the electorate's own general view of the government's performance. Monitoring elected agents is difficult and time consuming: time spent attending to politics is time not spent on something else. Thus the relative degree of monitoring presumably depends on the tradeoff between the perceived utility of paying attention to the government and letting government agents go unchecked.
When the electorate has reason to trust that government officials are acting in ways that make the public better off, the electorate has an incentive to spend less of its valuable attention on government. However, when there is evidence of wrongdoing, incompetence and/or ineffectiveness, the electorate has a stronger incentive to continually monitor elected officials at a perhaps lower but continuous cost. Essentially, when there is information that something has gone wrong or the electorate feels that the government is not trustworthy or is performing poorly, then voters should be more likely to attend to the actions of elected officials. When there is evidence of success, by contrast, voters should not waste their energies focusing on the activities of their representatives. The motivation for this could be cognitive: voters perceive a problem and are seeking information about a solution. It could also be a response to fear or anger as citizens worry about the fate of the nation even if there are not good solutions. In fact, it is probably some combination of these effects that produces the shifts in attention.
We argue that this concept of monitoring is best thought of as the electorate's aggregate level of political interest. Consistent with this reasoning, as we detail below, the most frequently used measures of political interest are not, curiously, 'how interested are you in politics?' Instead, interest is more often than not assessed by asking individuals how closely or how often they follow or pay attention to government, national politics, campaigns and the like. 'Following' and 'paying attention to' are actions in which someone who is auditing government affairs would engage. In sum, the most commonly used measures of political interest are analogous to the extent to which the public is monitoring the actions of its elected officials. 4 Understanding what features of politics are likely to lead to shifts in macrointerest requires a return to this notion that 'interest' in politics is best conceptualized as the extent to which someone is paying attention to what the government and political actors are doing (or the desire/motivation to pay attention). As such, any theory about the dynamics of macrointerest should begin with a discussion of the fundamental forces in American politics that lead people to be more or less likely to want to take time away from other (likely more rewarding) activities to turn their attention to the goings-on in government.

Hypotheses
The standard cross-sectional predictors of individual-level interest, such as education, income, race and gender, provide no theoretical leverage with regard to macrointerest. The proportion of the electorate that has a college degree or is female moves too slowly over time to be able to explain aggregate movements in interest. Instead, as with any work in American macropolitics, we need to focus on what creates small, but uniform, shifts in the macrolevel phenomenon.
We use the theory described above to derive several testable hypotheses about the specific conditions under which the polity is more or less likely to pay attention to the goings-on in government and governmental affairs. Consistent with Hetherington's (2005) finding that trust is positively associated with support for policies that expand the scope of government, we theorize that when people trust elected officials to behave as responsive agents, they will feel less of a need to expend their cognitive energies monitoring the goings-on in government. We argue that at least some of the handwringing about the decline in political trust is misguided. Rather than being a solely negative feature of American politics, declining political trust has a potentially positive effect. Specifically, our primary hypothesis is that changes in aggregate-level trust in government generate inverse changes in aggregate-level political interest. As the electorate becomes less trusting of government, the electorate will become more interested in politics, because it feels the need to monitor elected officials in order to find solutions to the problems at the root of the increasing distrust.
There are a few additional hypotheses that are consistent with our theory. Presidential job approval serves as a signal of overall governmental performance. If the electorate is more satisfied with the performance of the most visible agent, then there is less need to monitor politics, and interest should be lower as a result. More direct signals about government performance should also affect macrointerest. One indicator of government performance that has an impact on public opinion and vote choice is the state of the economy (for example, Kinder and Kiewiet 1979;Kinder and Kiewiet 1981, but see Anderson 2007). We hypothesize that when the public believes the economy is stronger (an indicator that elected officials are doing a good job), they will respond by becoming less interested in politics. In contrast, negative changes in perceptions of the economy will cue the electorate that the government is performing poorly and is in need of monitoring; interest should rise as a result.
In addition to public opinion and external indicators of government performance, we hypothesize that discrete negative events in modern American history should command the attention and interest of the electorate, either because of their severe negativity (for example, 9/11 or Hurricane Katrina and its aftermath) or because they are strong signals of government 4 It is important to qualify our assertion that political interest is a measure of the extent to which citizens are monitoring the actions of government officials in two ways. First, whether not citizens are effectively monitoring elected representatives (e.g., seeking out and integrating information about the goings-on of government in an unbiased fashion) is a separate question that is beyond the scope of this research. Secondly, our argument is a relative one. In other words, increases in political interest of the ordinary type (as measured in public opinion surveys) is probably not enough to truly qualify as monitoring in the normative sense. Rather, our focus here is on relative increases in the degree to which individuals are paying attention to the goings-on of government. We thank an anonymous reviewer for suggesting this qualification. malfeasance (for example, Watergate and the Keating Five banking scandal). We expect that macrointerest will be higher in the presence (vs. absence) of a scandal or major negative event.
In sum, based on our principal-agent theory of political interest, we derive our primary hypothesis about the cause of macropolitical interest: HYPOTHESIS 1: As the public's trust in government decreases, macrointerest will increase.
We also derive three related hypotheses: HYPOTHESIS 2: As aggregate presidential approval decreases, macrointerest will increase.
HYPOTHESIS 3: As the index of consumer sentiment decreases, macrointerest will increase. HYPOTHESIS 4: Macrointerest will increase in the presence (vs. absence) of government scandals or acutely negative events.

Alternative Theories
We also consider alternative explanations about which aspects of the political context stimulate changes in the electorate's level of political interest. These explanations are obviously not mutually exclusive; several specific hypotheses are connected to more than one of these theories.

Alternative Theory 1: The public does not care
One caricature of the public is that most Americans neither know nor care much about politics. From this perspective, the costs of interest outweigh the benefits, and therefore auditing the government and its policy outputs is just not worth the time. Also, if individuals believe that others are monitoring the activities of elected officials, they might free-ride, and leave the auditing to others who are probably free-riding themselves. The prediction derived from this alternative theory is that the level of aggregate political interest in the electorate will be quite low.
Whereas this first alternate theory focuses on the levels of interest, it also has important implications for the dynamics of changes in macrointerest. If the public does not care about politics and does not pay attention, any movements in aggregate interest should be statistical noise, because for the public to respond to changes in the political culture/context, it would have to know and care about those changes. If the macropolity is uninterested, it will miss these cues, which should result in a long-term equilibrium that is low in magnitude and, more importantly, decays back to the previous equilibrium quickly even after events, crises or shocks to the political system.
The opposing argument that interest should be mean reverting is relatively persuasive at first glance. If some people are simply interested in politics and some are not (as Prior 2018 finds), then this might imply a long-run equilibrium, and that any deviations from this mean level of interest should dissipate and the series should revert to its long-term mean. Why would we be skeptical about this? We argue that if microlevel changes in interest in politics are systematic and essentially permanent, then the aggregated series should not be mean reverting. It should have a unit root If individual-level changes are rare but permanent, then if some exogenous force were to increase the public's interest by, say, 5 per cent, we would not expect that effect to decay back to the prior mean. Instead, we would expect the series to stay at this new level until some other shock caused it to move to a new equilibrium. Therefore, a second alternative hypothesis derived from ALT1 is that not only will macrointerest be low, but it should be relatively stable and, as a time series, it should have a very short memory.
Alternative Theory 2: Politics is interesting when things are going well in the country Stimson (2015) makes the case that much of the over-time variation in public opinion across a range of macrolevel measures is really just an indication of the public's satisfaction with the government. For example, the healthier the economy, the higher the levels of presidential approval, congressional approval and trust in government. It is plausible that political interest is just another manifestation of this general pattern. This suggestion predicts exactly the opposite of the principal-agent theorythat macrointerest is positively related to presidential approval, trust in government and the index of consumer sentiment and negatively associated with the presence of scandals. If this theory is correct, and macrointerest is merely another indicator of positive perceptions of government performance, then there would not be much use for a measure of macrointerest.
Alternative Theory 3: The public becomes more interested in politics during a presidential election Interest in politics may also be driven by the content of politics. There is no higher drama in American politics than presidential elections. Given the stakes involved and the ubiquity of media coverage of these elections, it is relatively easy to imagine that presidential election campaigns will motivate the public to become more interested in politics. To be clear, the measurement of macrointerest will be based on public opinion survey questions that do not reference elections. In fact, some of the questions upon which we rely explicitly ask respondents how much they are interested in politics in general, not specifically during elections. Still, a presidential election has the potential to attract the attention of the public and heighten the electorate's interest in politics more generally.
Although this may seem relatively obvious at first, there is some evidence that this is not the case. Prior (2010Prior ( , 2018, for instance, suggests that elections have no effect on interest in politics at the individual level. Interest in elections, specifically, tends to respond to the episodic nature of American elections, but general interest in politics, he argues, is a different and stable force. It does not wax and wane along with the electoral cycle at the micro level.
To summarize, the alternative hypotheses are: ALT1, HYPOTHESIS 1: Macrointerest will be uniformly low.
ALT1, HYPOTHESIS 2: Macrointerest will quickly return to its long-run equilibrium.
ALT2, HYPOTHESIS 1: As the public's trust in government increases, macrointerest will increase. ALT2, HYPOTHESIS 2: As aggregate presidential approval increases, macrointerest will increase. ALT2, HYPOTHESIS 3: As the index of consumer sentiment increases, macrointerest will increase.
ALT2, HYPOTHESIS 4: Macrointerest will increase in the absence (vs. presence) of government scandals or acutely negative events.

Developing a Measure of Macrointerest
In order to examine the extent and causes of variability in aggregate political interest (that is, macrointerest), we need a measure of the public's interest in politics over time. The dependent variable is the electorate's level of interest in politics over time. A measure of this concept has not yet been developed at the macro level. The most obvious approach to constructing such a measure would be to use the cumulative ANES. In most years of the study, the ANES asks respondents: 'Some people seem to follow what's going on in government and public affairs most of the time, whether there's an election going on or not. Others aren't that interested. Would you say you follow what's going on in government and public affairs most of the time, some of the time, only now and then, or hardly at all?' Figure 1 plots the percentage of respondents in the top two categories of this survey question for the duration of the cumulative ANES. 5 We argue that this measure is inadequate because it suffers from three serious issues. First, the data are biennial and do not include every even year. It is therefore difficult to use these data for time-series work because the uneven spacing of the series dramatically complicates the estimation of the dynamics. Secondly, the measure is based on a single question wording. Although the ANES wording has become the industry standard, it is not the only measure of political interest that has been used in the literature (see, for example, Prior 2010). Thirdly, the ANES is a deeply political survey conducted during the heart of an election year (either presidential or off-year). If either the broader political context or the specific content of the other political questions on the survey primes responses to the interest question (thus biasing expressions of interest upward), then it is possible that the ANES understates over-time variability in political interest.
Pew has also incorporated the same question into fifty-six of its surveys since the 1990s. Whereas the question wording is the same, not all of the surveys in which Pew asked the interest question were fielded during election years or were explicitly about politics. Figure 2 plots the movement in the interest question in these Pew surveys over the same timeframe. This survey shows much more movement than the ANES. The one problem with the Pew surveys for timeseries work is that they do not start until 1990. Relying on only the Pew data would limit the timeframe of the analyses.
There are still other options. Beginning in the 1970s Roper asked its respondents 'Would you say that you have recently been taking a good deal of interest in current events and what's happening in the world today, some interest, or not very much interest?' This question, which was asked 107 times, is likely to tap similar considerations in respondents and gets at the general concept of interest in politics. Figure 3 presents the movement in this series. Like the Pew series, the Roper interest in politics series exhibits more movement than the ANES series, but also has a shorter timeframe than the ANES.
An ideal measure of macrointerest would combine all of these measures of political interest, and any others available, into a single measure that would: (1) cover a long timeframe, (2) vary at a smaller level of temporal aggregation than biennially and (3) capture the shared variance in the dynamics of the individual series. This assumes, of course, that there is common variance between each of these indicators that could be explained. To help visualize whether there are similar movements in the plotted series, we standardize the series by subtracting off the mean of each series and then dividing by the series' standard deviation, add one from Gallup and a second series from the Pew surveys (see Appendix A for a full list of the question wordings used in these analyses), and put them into a single figure (Figure 4). To be clear, this is only to aid in the visualization of the series; the unstandardized series are used in the actual measurement scheme. As is evident from the graph, the ANES series seems to have much less variability than the other series, but it has the longest timeframe. Each of the series, however, has a fair amount of movement over time.
Simply visualizing the series is useful, but it does not provide any definitive evidence of the commonality between the series. To capture this shared variance and create a measure of macrointerest, we employ the dyad ratio algorithm technique used by Stimson (2018)  The dyad ratio algorithm (more commonly known as the software Wcalc) takes disparate survey items that generally measure the same concept, and finds the shared over-time variance. The basic idea is that if there are multiple survey items that are each asked at several points in time, the movement over time in each of the items will be a function of both the idiosyncratic variance in each item and the shared variance in the movement of the underlying concept in the aggregate population. The key to the measurement scheme is to have multiple, overlapping items. It is not necessary for each survey item to be asked at each time point. The Wcalc procedure uses the ratio of the different items at the same points in time, when available, to determine how the items vary similarly. From this, the algorithm extracts the common variance and generates a single series that summarizes a quarterly measure of the electorate's interest in politics (Stimson 1998;Stimson 2015;Stimson 2018).
The key to the algorithm is including as many survey items that measure the same basic concept as possible. Figure 4 presents five of these possible series, but they are not the only measures  Stimson's (2018) process for developing his measure of public policy mood, we searched the Roper database of survey questions to find any survey that asked the public about its interest in politics. We began by entering a set of keywords into The Roper Center's iPoll, limiting the search to national US polls. We used three search terms: interest, attention and follow. We omitted the questions that asked specifically about interest in an upcoming or recent election and those that were clearly about a different topic ('interest groups'). We also added the standard political interest question from the cumulative file of the ANES to these data. We found 329 surveys that asked some form of the political interest question. These 329 surveys came from forty-six series made up of unique combinations of survey firm and question wording. In Figure 4, for instance, we include two different series of surveys from Pew. The first of these asks the standard ANES-based question. The other asks, 'Generally speaking, how much interest would you say you have in politics…a lot of interest, some interest, only a little interest, or no interest at all?' Because they have different question wordings, Wcalc treats them as two different series.  To be included in the macrointerest measure, the Wcalc procedure requires that each specific series (that is, combination of question wording and survey firm) occurs at more than one point in time. Of the forty-six different series included in the data, only sixteen occurred in more than one quarter. 6 These sixteen series are the raw survey marginals included in the analysis. The data range from the first quarter of 1973 to the first quarter of 2014.
The Wcalc procedure finds the common movements over time in variables. The underlying idea is that the public has a latent opinion on the variable (in this case interest) that moves over time. Variance in political interest, as measured by the same survey firm across multiple points in time, is a function of: (1) latent underlying public opinion and (2) idiosyncratic features of the survey firm, which includes question wording and 'house effects' over time. Moreover, because the survey questions are asked at various points in time and no single series is asked regularly enough over the entire range of the combined time series, no single series can be used for every time point.
Essentially, Wcalc solves both of these problems simultaneously by imputing the values of the missing series for each time point for which it is not observed based on the estimated ratio of the missing survey items to the observed survey items from that time period. For example, if the aggregate value of political interest in the ANES series is, on average, equal to 0.90 of the value of the aggregate political interest measure in the Pew series, then Wcalc imputes the value of the ANES political interest measure for the time period in which it was not measured by calculating the value of the Pew survey political interest measure (for that time period) * 0.90. Wcalc does this simultaneously for all of the series and finds a single latent dimension estimate, which is our measure of macrointerest. The value at a specific time point can be thought of as the percent of the public that would respond that he or she was interested in politics on a typical survey item (Stimson 2018).
One other issue must be addressed before we can construct a measure of macrointerest. The Wcalc procedure is designed to deal with the ratio of survey respondents who do or do not hold certain attitudes. In the measurement of policy mood, this is operationalized as the ratio of left-leaning respondents divided by the sum of left-and right-leaning respondents. Those who report a moderate opinion are omitted from this calculation. Similarly, in Stimson's work using this technique for measuring macropartisanship, the ratios are the number of Democrats divided by the sum of Democrats and Republicans, omitting independents.
The coding method we used for the political interest items is consistent with Stimson's approach. If the survey had three responses, the input is the category indicating the 'most' political interest divided by the sum of the most and the least (omitting the middle category, don't knows and refusals). If there are four categories, the input is the sum of the highest two categories divided by the sum of the four categories (omitting don't knows and refusals). If there are five categories, the middle category, don't knows and refusals are omitted; the inputted marginal is the highest two categories divided by the sum of the two highest and two lowest. The best example of this is the question asked by Roper quoted above, which gives respondents three categories: a good deal of interest, some interest and not very much interest. In this case, our input to the Wcalc procedure is the ratio of respondents who answer 'a good deal' divided by the sum of those who say 'a good deal' and 'not very much', omitting the 'some interest' respondents. Only two other series (CBS News and AP/gfk Knowledge Networks) have an odd number of response categories that require this measurement decision. For all of the series, we omit the don't knows and refusals. 7 Table 1 lists the survey firms, the number of times that firm asked the same question (question wording available in Appendix A) and the loading for the series. Gallup and Princeton Survey Research Associates (PSRA) have multiple entries because they ask different questions. These loadings are essentially the over-time correlation between the items and the political interest latent variable. Relative to the loadings reported in several other uses of Wcalc, these are all quite high. 8 For instance, Stimson (2018) analyzes various polling firms' measurement approval of President Trump. In those results, two of the firms (Quinnipiac and CBS/New York Times) have loadings below 0.3. Having all the lowest of these loadings just below 0.5 is remarkably high for these types of data. Additionally, the results suggest that the latent macrointerest measure captures a sizable proportion of the variance in the individual series -66.5 per cent. There is no evidence of a second dimension in the data, and the latent variable explains about two-thirds of the variance in the individual series. This is about double the typical variance explained for something as complex as policy mood, but somewhat less than the variance explained by comparable metrics for presidential approval (Stimson 2018). Figure 5 returns to the plot of the five series from Figure 4 and adds the measure of macrointerest (the black line). 9 The macrointerest measure is generally in the middle of the component series, and the common latent variable explains over two-thirds of the variance in the individual items. Macrointerest is more unstable than the ANES series, but is more stable than the other four. As Figure 5 shows, the electorate's interest in politics moves substantially over the time period for which we have data. Its nadir (65.4) was in the mid-1980s and its peak was in the first quarter of 2013 (83.8). There does not appear to be any obvious periodicity in the series, as we would expect if ebbs and flows in macrointerest were simply driven by the timing of elections, though we will test for this empirically.
One other possible concern is that the movement in the series is simply the result of stochastic noise or measurement error. We do not think this is a serious concern. The Wcalc procedure The negative loading for the RT Strategies polls is not uncommon for items with only two time points (Stimson 2018). All it means is that while the rest of the series is moving in one direction at those two time points, these two time points move in the opposite direction. Because the items fit our initial criteria for inclusion of the survey, we follow the best practices with this approach and leave these items in the analysis. finds the common variance between the multiple items. Instead of relying on a single set of survey items, which would raise some of these concerns, the measure we use detects the common dynamic variance, which should dramatically decrease this concern.
To determine whether the series is dependent on a single survey in the data, we re-estimated the macrointerest measure multiple times, each time with one of the surveys removed. In every case, the correlation between the measure that includes the full sample and the one with a series removed is greater than 0.80 and most of them are higher than 0.90. Our interpretation of these results is that (1) no single survey is dominating the results and (2) there is not a lot of noise in the data. If there were a lot of noise, we would expect the series to move substantially with the inclusion or exclusion of individual items. The fact that we find an incredible amount of consistency suggests that the macrointerest measure is mostly signal and little noise. Figure 6 plots the macrointerest series on its natural metric instead of the standardized version presented in Figure 5 and includes a line summarizing how many surveys there are in each quarter. 10 To capture the uncertainty estimate in the measure we followed the procedure outlined in Stimson (2018) and calculated the standard errors of the estimate via a bootstrap. For most of the time points, the uncertainty in the estimates is quite small; 90 per cent of the time points are estimated as having a standard error of less than half a percent. The few time points that have noticeably large standard errors (for instance, the first two quarters of 2002 and the last quarter of 2005) are due to differences in the movements of the component series. The last quarter of 2005, for instance, has an RT strategies poll. Because it has a negative factor loading (see Table 1) it increases the uncertainty in the estimate for that time point.
We can compare the amount of movement in this series to that of other familiar series. Stimson (2015) uses the same techniques to create aggregate measures of presidential approval, macropartisanship and domestic policy mood. Not surprisingly, presidential approval has easily the widest range. It bottoms out during Watergate (28.5) and reaches its peak right after 9/11 (90.2). The 18.5-point movement in macrointerest pales in comparison and is more akin to the movements in policy mood and macropartisanship. In the same timeframe we examine, macropartisanship only moves around 13.5 points and domestic policy mood has a range of 15.5 points. In summary, macrointerest may not be as wide ranging as presidential approval, but it appears to have similar movements to several other important macro time series. Most importantly, this basic finding runs contrary to prior thinking about the stability of aggregate interest over time.

Independent Variables
To test the predictions derived from our theorythat levels of macrointerest will rise when the public believes things in the country are not going wellas well as the opposite predictions derived from ALT2 (that macrointerest will rise when the public believes things in the country are going well), we gathered aggregate measures of trust in government, presidential approval, consumer sentiment, and the presence of major scandals and acutely negative events.

Trust in the Federal Government
The quarterly measure of trust follows the lead of Keele (2005), who combined all survey recordings of trust in the federal government via the dyad ratio algorithm. These data were updated and provided by Nakajo (2015).

Presidential Approval
The presidential approval series used here was developed by Stimson (2015). It combines all available recordings of presidential approval via Wcalc.

Index of Consumer Sentiment
The measure of the perception of the economy is the standard Index of Consumer Sentiment, created by the University of Michigan Survey of Consumers (http://www.sca.isr.umich.edu/ tables.html). This captures the public's perception of the state of the economy and tends to be a stronger predictor of political behavior and attitudes than actual measures of the economy (Erikson, MacKuen and Stimson 2002). 11

Scandals and Negative Events
We include the relatively standard collection of indicators of major scandals and acutely negative events in American politics during our timeframe: Watergate, ABSCAM, the resignation of Speaker Wright, the Keating 5, the impeachment of Bill Clinton, September 11 and Hurricane Katrina. We also control for major military events in our timeframe including the invasion of Panama, the first Gulf War and the invasion of Iraq. 12

Presence of Presidential Election Cycle
Given that presidential elections are the most salient elections in the US context, as demonstrated by their relatively high participation rate, we include a variable that is coded 0 for all quarters that are not in a presidential election year, and counts the quarter in the presidential election year. The first quarter of 2000, for instance, is scored as 1, the second is 2 and so on. All four quarters of each non-presidential election year are scored 0. 13

Methods
The choice of models with time series can be complicated. 14 Based on our analysis, the results indicate that our interest series and our trust series are integrated. Therefore we estimate the relationship as an error correction model (ECM). The ECM has a particular structure to it that many now find familiar: where the first difference of y t is specified as a constant term (α 0 ), the lagged value of y t , both the change in and lagged value of x t and an error term ϵ t . The coefficient β 0 summarizes the effect of a contemporaneous change in the independent variable at time t on changes in the dependent variable. β 1 represents how the lagged level of the independent variable changes the dependent variable. The α 1 term on the lagged value of y t captures the rate at which the two series (interest and trust in our case) return to equilibrium after a shock. There is a more intuitive interpretation of the ECM. Our primary research question is: what is the over-time relationship between trust and interest? There are two ways that trust can be related to interest. First, a change in the level of trust in government may be related to an immediate change in interest. This is captured by the β 0 term. This is important, but it does not get at the long-run equilibrium between trust and interest. The second type of over-time relationship involves the cointegrating relationships. If something causes trust to decline and be out of equilibrium with interest, how does this effect create a longrun change in interest, one that builds over time as long as trust and interest are out of 11 We have used several measures of the actual economy as opposed to perceptions of the economy as well. Unemployment, inflation, change in per capita GDP and real disposable income all result in the same empirical patterns as the Index of Consumer Sentiment. 12 Including the invasion of Afghanistan is essentially the equivalent of including 9/11. 13 We have tried several other specifications to capture the effect of the election cycle and the results do not change.
14 In Appendix B we discuss our choices in more detail, providing the results of several specification tests. Our model is based on the Engle-Granger tests and the autoregressive distribute lag-bounds method suggested by (Phillips 2017). Our findings indicate that there is a co-integrating relationship between the two series. Additionally, we find evidence that trust Granger causes interest and that interest does not Granger cause trust.
equilibrium? The answer to this is captured, in part, by the β 1 term and is substantively the more important piece of the ECM.

Results
We can evaluate the first alternative theory directly. The stability theory implies that the interest series should have a stable equilibrium. Whereas the series may move from random events or shocks, these effects should decay quickly. The unit root tests indicate that this is not the case. The test statistic for the augmented Dickey-Fuller test is −0.07, which is well below the threshold to reject the null hypothesis of a unit root (see Appendix B). An integrated series, by definition, simply does not exhibit these features. A shock to the macrointerest series does not decay, but creates a permanent change. The co-integrating relationship between trust and interest does have a long-run equilibrium and shocks to that equilibrium will decay, but the stationarity tests (see Appendix B) suggest that macrointerest does not have a long-run equilibrium. The results of the diagnostic tests, then, lead to the rejection of ALT1 H1 and ALT1 H2. Macrointerest does not have a short memory, and it does not have a set equilibrium.
We wish to reiterate that finding that interest is integrated and not mean reverting is entirely consistent with the microlevel stability results. Changes in interest at the individual level show little sign of decaying. Once someone becomes interested in politics, they tend to stay interested (Prior 2018). The aggregate-level consequence, then, should be that the macrolevel interest series is integrated, which is exactly what we find here. Changes in interest are permanent and do not decay. The current level of interest is not just a function of recent events, but the accumulation of all past events. When an event or moment makes the public more interested in politics, these effects do not dissipate. There is no long-run equilibrium level of interest that is natural for the public. Instead, macrointerest movesand when it moves, it stays until some other forces cause it to change again. 15 Before turning to our multivariate model, we explore the bivariate relationship between our dependent variable, macrointerest, and the three other series that we hypothesize will be related to macrointerest: trust in government, the index of consumer sentiment and presidential approval. Figure 7 depicts the relationships between macrointerest and the three other series. To highlight the common movements over time, we have standardized each of these series to make them more comparable visually; all of the statistical analyses that follow rely on the original scaling of the variables. In each panel, macrointerest is plotted as a solid black line and the independent variable is the gray dashed line. In each case, there is a negative relationship between the two variables. When the independent variable is higher, macrointerest is lower. The simple correlations between the series suggest that this pattern is strongest for trust in government (−0.55) and weaker for presidential approval (−0.41) and the index of consumer sentiment (−0.21).
One of the strengths of ECMs is that they allow researchers to separate out the immediate effect of changes in the independent variables, captured by the coefficients, from changes in the independent variables from the long-run effects, captured by the coefficients on the lagged value of the independent variables. In addition, because of the dynamics in the model, researchers can determine the cumulative effect of a change in the independent variable by calculating the long-run multiplier. Table 2 presents the results from the ECM predicting changes in macrointerest. 16 The first parameter of interest is the coefficient on the lagged dependent variable. When variables are 15 It is important to note that there is no yardstick to help a researcher determine whether the variability evidenced in a measure of a construct is 'a lot' or 'a little'. We argue that the diagnostics we report here make a convincing case for rejecting ALT 1, H2 (that macrointerest is relatively stable and has a short memory). There is also no benchmark to determine how much variability in the measure of a construct is systematic or random (e.g., due to statistical noise in the original data or modeling noise). We argue that because (1) the macrointerest series is not dependent on a single survey in the data and (2) the series is integrated, at least some of the variance in macrointerest is systematic. 16 The Ljung-Box test of the residuals from this model indicates that they are white noise.
co-integrated, like trust and macrointerest, they have a long-run equilibrium. In these models, this coefficient captures the speed at which the two series move back towards equilibrium when they become out of phase. This error correction rate (the lagged coefficient plus one) indicates that if Figure 7. Bivariate relationships between macrointerest (in bold) and trust, presidential approval, and consumer sentiment the series are out of equilibrium by one unit, at time t, they will be out of equilibrium by approximately 0.86 units in t + 1. The half-life of this effect is five quarters. After that point, a one-unit shift at time t has less than a 0.5 effect on macrointerest. This is a relatively slow movement back to the long-run equilibrium between trust and interest. Whereas the error correction results imply that when macrointerest is shocked out of equilibrium it will return to where it should be based on trust in government, its return is relatively slow. Trust has both an immediate and a long-run negative effect on macrointerest. The coefficient for the change in the trust variable indicates that a one-unit increase in trust at time t produces an immediate 0.21-point decline in macrointerest. The effect of this same change in trust also accumulates over time. The coefficient on the lagged value of trust indicates that if the level of trust is one unit higher, macrointerest is expected to change by −0.09 units in the next quarter. This effect, however, decays slowly and accumulates via the error correction rate. The long-run multiplier (calculated as the ratio of the coefficient for the lagged effect of trust divided by one minus the error correction rate) captures this effect. For trust, the long-run multiplier is −0.63 (with a confidence interval of −0.17, −1.08), 17 meaning that the cumulative effect of a one-unit shift in trust, captured by both of the coefficients, is about three-fifths of a point decline in the public's interest in politics. The mean absolute change in interest is 0.99 points from one quarter to the next. A one-unit change in trust thus produces about two-thirds of an average change in interest.
One potential concern with these results is that they are based on some particularly influential time points. Looking at Figure 7, for instance, both the beginning and end of the macrointerest series have particularly high values. It is possible that one or both of these eras may be exceedingly influential on the results. However, omitting either the first or last two years of the series (or both) does not alter the conclusions. The inverse relationship between trust and interest is robust to these specifications. 18 One of the events is also a significant predictor of changes in macrointerest. Consistent with our Hypothesis 4 and inconsistent with ALT2 Hypothesis 4, September 11, probably the most consequential event in the timeframe, has a positive and significant effect on changes in interest. In the third quarter of 2001, there was a 1.90 change in interest beyond what would have been expected from the rest of the model. Inconsistent with Hypotheses 2 and 3, and with ALT2 Hypothesis 2 and ALT2 Hypothesis 3, neither of the other independent variables has a significant effect on macrointerest. Despite the bivariate correlations displayed earlier, there is no evidence that presidential approval or the index of consumer sentiment has either an immediate or longrun effect on macrointerest in the full model. Tests of the joint significance of both coefficients for each independent variable are also insignificant.
The episodic moments in American politics can affect the macrointerest series. Consistent with ALT3 Hypothesis 1, the presidential election variable is a significant predictor of changes in macrointerest. In presidential election years, the country becomes increasingly interested in politics. Keep in mind that none of the constituent series in the macrointerest measure assesses interest in the election specifically: they instead assess general interest in politics and government. Some of the indicators, in fact, explicitly instruct respondents not to think about the election. Still, the electorate is increasingly more interested in these years than in others. None of the other events has a significant effect on changes in macrointerest. 17 Calculated via the Delta method. 18 As an additional robustness check, we re-estimated the error correction model reported here, each time removing one of the surveys that comprises the macrointerest measure. With two exceptions, removing a single survey does not change the substantive conclusions. The two exceptions are the ANES series and the Roper series (omitting either one of these reduces the effect of trust on interest). It is not surprising that these are the two most important series, because the Roper series has the most time points and the ANES has the longest timespan. Our read of the results is that omitting either of these series substantially decreases the signal-to-noise ratio in the macrointerest series, leading to more measurement error and less robust results.

Discussion
What makes a good, or at least better, democracy? Certainly, there are institutional features such as an independent judiciary and free and fair regular elections, but attributes of the electorate matter as well. All else equal, a democracy in which the public is interested in politics is preferable to one in which it is not. Similarly, trust in government makes democracies function more effectively. The paradox our results present is that interest and trust are negatively correlated. When the electorate trusts government more, people tend to become less interested in government. In contrast, when something leads the public to become less trusting in government, the result is greater attention to the activities of the government.
A normatively displeasing interpretation of these findings is that they signify a relatively tragic failure of the American system (despite the upward trend in macrointerest over the last two-plus decades demonstrated in Figure 6). The finding that macrointerest responds positively to negative circumstances runs counter to the individual-level findings that high levels of interest are associated with positive outcomes (such as increases in political participation, for example, Verba, Schlozman and Brady 1995), and the presumption that high levels of interest are therefore the bellwether of a healthy democracy and good times. Combining these macrointerest findings with the microlevel findings that political interest is arguably the most consistent and most powerful predictor of political participation, and with all the appropriate caveats about ecological fallacy, it is possible that the increase in participation we see as a result of increases in political interest at the micro level can be attributed to what we might call 'negative participation'.
These findings add to the literature that points to the importance of trust in government. At its most basic level, political trust is associated with the support necessary for a healthy representative democracy; as trust decreases, support for more direct democracy increases (Karp 1995; see also Hetherington 2005) and compliance with laws decreases (Scholz and Lubell 1998; mistrust has also been causally linked to support for conservative domestic policies (Hetherington 2005)). Our work contributes to this literature by showing that political mistrust also serves as a fire alarm that leads to an increase in macrointerest, which we argue is an essential prerequisite to the public's motivation to monitor the goings-on in government and to keep elected officials in check.
On the one hand, our findings could be interpreted as pointing to a potentially positive side to declines in trust, painting a more positive picture of American politics. When the electorate becomes less trustful, it responds as any reasonable principal should: it becomes more interested, and therefore more motivated to monitor its principals. On the other hand, our findings could be interpreted as challenging the conceptual link between political interest and democratic health. To the extent that macrointerest is inversely related to the public's trust in government (itself an indicator of democratic health), then increases in interest at the macro level could indicate a decline in democratic health. We look forward to cross-national replications that adjudicate between the 'glass half full' and 'glass half empty' interpretations of our findings.
There are also several possible extensions to these results. One clear one would be to test whether there are differences in the levels of interest between members of the president's party and out-partisans. Our entire focus, as is the case with macrolevel work, is on the overall patterns of the full electorate. Splitting the macrolevel measures based on some feature of the electorate is not impossible (see, for instance, Enns and Kellstedt 2008), but it does create limitations on the measurement. The vast majority of the indicators that make up the dependent variable are only available as the marginals from a national survey; they do not provide the breakdown by party (or other potentially interesting splits). Testing for these 'meso' level patterns is an important next step, but requires a different design.
Our results are suggestive of the nature of macropolitics in America more generally. Given that most Americans are not particularly politically astute or organized (Delli Carpini and Keeter 1996;Olson 1965), it is not clear why we should expect the government to ever heed the unwashed masses. Yet even the most skeptical observers of American government would not view the people as powerless, nor the elites as completely unconstrained. We argue that interest is a precursor to voice and participation: if the public is not paying attention, it will not know when its voice is most needed. Because interest in politics is such a fundamental commodity in a representative democracy, the systematic variation in macrointerest shines light on the conditions under which the polity actively participates in its own representation by monitoring and holding its agents accountable for their performance. The basic nature of dynamic representation and the link between the public's preferences and the policies the government enacts may very well depend on the levels of interest in the electorate.
In sum, we find that the electorate becomes more interested and attentive to government when times are bad (in particular, during times of low trust in government), but is otherwise willing to let its unpopular agents off leash (possibly counting on other linkages, such as the media, interest groups and other actors to sound the 'fire alarms' when necessary). In other words, our findings paint a picture of the American polity as being an increasingly attentive 'god of vengeance' and punishment but a lazy 'god of reward' (Key 1966