Cross-National Macrointerest: The Source Data

National and cross-national surveys have asked questions on political interest, often over the past four decades, but the resulting data are both sparse, that is, unavailable for many countries and years, and incomparable, generated by many different survey items. In all, fifty-four such survey items were asked in no fewer than five country-years in countries surveyed at least twice; these items were drawn from 396 different survey datasets (see online supplementary appendix).

Together, the survey items in the source data were asked in 127 different countries in at least three time points over the forty-one years from 1982 to 2023, yielding a total of 2,738 country-year-item observations. Observations for every year in each country surveyed would number 5,207, and a complete set of country-year-items would encompass 281,178 observations. Compared to this hypothetical complete set of country-year-items, the available data are very, very sparse. More optimistically, there are 1,821 country-years in which there is at least some information about the public’s interest in politics, that is, some 57 per cent of the 3,220 country-years spanned by these data. Still, the multitude of different survey items makes these data incomparable and difficult to use together.

In the top left panel of Figure 1, the twelve countries with the most country-year-item observations are displayed. Germany, with 122 observations, is the best-represented country in these source data, followed by the UK, Spain, the USA, and the Netherlands. At the other end of the scale, there are seven countries – Azerbaijan, Cambodia, Kosovo, Kyrgyzstan, Liberia, Myanmar (Burma), and Puerto Rico – that have only the bare minimum of three observations needed to be included in the source dataset at all. In the top right panel are the dozen countries with the most observed years; this group is similar to that on the left, but with Ireland and Greece adding to the list and Finland and Sweden dropping off. The bottom panel shows the number of countries observed in each year. Coverage across countries reached its apex in 2008 when respondents in eighty-nine countries were asked at least one item about their interest in politics. The next section describes how this sparse and incomparable survey data was used together with a latent variable model to generate a complete time series of macrointerest scores that are comparable across countries.

Figure 1. Countries and Years with the Most Observations in the Source Data.

Estimating Cross-National Macrointerest

Several recent studies have developed latent variable models of aggregate survey responses based on cross-national survey data (see Claassen Reference Claassen2019; Caughey, O’Grady and Warshaw Reference Caughey, O’Grady and Warshaw2019; McGann, Dellepiane-Avellaneda and Bartle Reference McGann, Dellepiane-Avellaneda and Bartle2019; Kolczynska et al. Reference Kolczynska, Bürkner, Kennedy and Vehtari2020). To estimate the public’s interest in politics across countries and over time, this work employs the latest of these methods that is appropriate for data that are both incomparable and sparse, the Dynamic Comparative Public Opinion (DCPO) model elaborated in Solt (Reference Solt2020b). Solt (Reference Solt2020b) demonstrates that the DCPO model provides a better fit to survey data than the models put forward by Claassen (Reference Claassen2019) or Caughey, O’Grady and Warshaw (Reference Caughey, O’Grady and Warshaw2019). The McGann, Dellepiane-Avellaneda and Bartle (Reference McGann, Dellepiane-Avellaneda and Bartle2019) model depends on dense survey data, unlike the sparse data on interest in politics described in the preceding section. Kolczynska et al. (Reference Kolczynska, Bürkner, Kennedy and Vehtari2020) is the very most recent of these five works and builds on each of the others, but the MRP approach developed in that piece is suitable not only when the available survey data are dense but also when ancillary data on population characteristics are available, so it is similarly inappropriate to this application. The dyad ratio algorithm employed in Peterson et al. (Reference Peterson, Miller, Saunders and McClurg2022), of course, leverages only over time variation within a single country and not variation across countries, making it a poor choice for generating cross-national estimates (see Caughey, O’Grady and Warshaw Reference Caughey, O’Grady and Warshaw2019, 686).Footnote ¹ The DCPO model is a population-level two-parameter ordinal logistic item response theory model with country-specific item-bias terms. For a comprehensive description of the DCPO model, see the supplementary appendix and Solt (Reference Solt2020b, 3–8); the focus here is on how it deals with the two principal issues raised by the source data, incomparability and sparsity.

The DCPO model accounts for incomparability using three sets of parameters. First, it incorporates the difficulty of each question’s responses, that is, how much interest in politics is indicated by a given response. This is most evident with respect to response categories: to say that one is ‘very interested’ in politics, for example, is to exhibit more interest than to say that one is ‘somewhat interested’ or ‘not very interested’. Here, difficulty is permitted to vary with question wording and the survey project as well. Second, the DCPO model accounts for each question’s dispersion, its noisiness with regard to our latent trait. The lower the dispersion, the better that changes in responses to the question map onto changes in macrointerest. Third, to provide for the possibility that translation issues or cultural differences result in the same question being interpreted differently in different countries, the model estimates country-specific bias parameters that shift the difficulty of all responses for a particular question in a particular country. Together, the model’s difficulty, dispersion, and country-specific bias parameters work to generate comparable estimates of the latent variable of macrointerest from the available but incomparable source data.Footnote ²

To address sparsity in the source data – unpolled or thinly surveyed years in each country – DCPO uses simple local-level dynamic linear models, that is, random-walk priors, for each country. That is, within each country, each year’s value of macrointerest is modelled as the previous year’s estimate plus a random shock. These dynamic models smooth the estimates of macrointerest over time and allow estimation even in years for which little or no survey data is available, albeit at the expense of greater measurement uncertainty.

The model was estimated using the DCPOtools package for R (Solt, Hu and Tai Reference Solt, Hu and Tai2019), running four chains for 1,000 iterations each and discarding the first half as warmup, leaving 2,000 samples. The $\hat R$ diagnostic had a maximum value of 1.01, indicating that the model converged. The dispersion parameters of the survey items indicate that all of them load well on the latent variable (see supplementary appendix). The result is estimates, in all 3,220 country-years spanned by the source data, of the mean political interest of the public, that is, macrointerest.

Validating Cross-National Macrointerest

That we can generate estimates of macrointerest does not automatically mean that they are suitable for analysis. As is the case for any new measure, validation tests of cross-national latent variables are crucially important (see, for example, Hu et al. Reference Hu, Tai, Ko, Woo and Solt2023). Figures 2 and 3 provide evidence of this measure’s validity with tests of convergent validation and construct validation. Convergent validation refers to tests of whether a measure is empirically associated with alternative indicators of the same concept (Adcock and Collier Reference Adcock and Collier2001, 540). In Figure 2, the macrointerest scores are compared to responses to individual source-data survey items that were used to generate them; this provides an ‘internal’ convergent validation test (for an example in a similar context, see Caughey, O’Grady and Warshaw Reference Caughey, O’Grady and Warshaw2019, 686).

Figure 2. Internal Convergent Validation: Correlations Between Macrointerest and Individual Source-Data Survey Items.

Figure 3. Construct Validation: Correlations Between Macrointerest and Other Aspects of Political Engagement.

On the left, macrointerest scores are plotted against the percentage of respondents across all country-years who offered the two most interested responses on the European Social Survey’s four-point item, ‘How interested are you in politics?’ The middle panel shows responses to the question with the most data-rich cross-section, ‘How interested would you say you personally are in politics?’ in the ISSP’s 2004 Citizenship module. Finally, the right panel evaluates how well the macrointerest scores capture change over time by focussing on the item with the largest number of observations for a single country in the source data, which asked respondents to Germany’s ALLBUS, ‘How interested in politics are you?’ In all three cases, the correlations, estimated taking into account the uncertainty in the measures, are strong.

Construct validation refers to demonstrating, for some other concept believed causally related to the concept a measure seeks to represent, that the measure is empirically associated with measures of that other concept (Adcock and Collier Reference Adcock and Collier2001, 542). Figure 3 depicts the relationships between macrointerest and three survey items from the WVS and EVS on other aspects of political engagement that are expected to have causal relationships with political interest (see Kittilson and Schwindt-Bayer Reference Kittilson and Schwindt-Bayer2010, 995): in the left panel, following political news on television, radio, and newspapers; in the center panel, discussing politics with friends; and on the right, feeling politics is important to one’s life. These relationships are all positive and are moderate to strong. This cross-national latent variable of macrointerest performs well in validation tests.

Testing Theories of Macrointerest Cross-Nationally

The best-developed theories of macrointerest concern the advanced democracies, and even among these relatively similar countries, macrointerest varies greatly. Figure 4 examines levels and trends in macrointerest in advanced democratic countries by displaying the changes of the public’s expressed interest in politics over time in the thirty-seven democracies of the OECD (supplementary appendix presents these macrointerest data for all available countries). While macrointerest scores approach and often exceed 0.6 in countries such as Denmark and Canada, in Chile, they scarcely cross 0.25. And although the public’s political interest has held fairly steady over decades in many countries, in Czechia it dropped nearly half of the variable’s entire theoretical range over the 1990s and 2000s before rebounding slightly since 2010, and increases of roughly a quarter of that range can be seen in, among others, Germany. There are considerable differences in the extent to which the public professes interest in politics both across countries and over time.

Figure 4. Macrointerest Scores Over Time Within OECD Democracies.

What accounts for these differences? The literature offers a range of arguments for how the political context may influence the public’s interest in politics. Perhaps the most straightforward is that publics grow more interested in politics at election time. Campaigns and elections attract media coverage and increase the information available to the public on the issues being contested, leading to increased interest in politics (see, for example, Larsen Reference Larsen2022). Macrointerest within a country should be expected to be higher, therefore, in years in which national elections take place than in years without elections.

A second argument is that political institutions that share power rather than concentrate it yield politics that are more interesting and engaging. Building on Lijphart (Reference Lijphart1999) and Powell (Reference Powell2000), Kittilson and Schwindt-Bayer (Reference Kittilson and Schwindt-Bayer2010, 992) argue that power-sharing institutions – parliamentarism, federalism, and proportional electoral rules – ‘send signals of inclusiveness to citizens, generating greater political engagement’ while power-concentrating institutions ‘may generate perceptions of exclusion and deter involvement’. Macrointerest should be higher in countries with parliamentary and federal systems than in those without those features, and it should decline as the disproportionality between votes cast and seats won increases.

A third claim deals with the public’s demand for accountability. Peterson et al. (Reference Peterson, Miller, Saunders and McClurg2022, 203) argue that ‘when there is information that something has gone wrong … then voters should be more likely to attend to the actions of elected officials’, but when ‘there is evidence of success … voters should not waste their energies’. If democracy is a principal-agent problem with elected officials acting as self-interested agents and the public as their lazy but vengeful principal, then macrointerest should rise when times are bad and decline as conditions improve.

A final set of theories – each well established – contradicts the third. Modernization theory holds that the public’s interest will increase as the national economy grows and household incomes expand (see, for example, Inglehart and Welzel Reference Inglehart and Welzel2005). Unemployment has long been argued not to motivate but to depress political interest (see, for example, Rosenstone Reference Rosenstone1982, 26). And the relative power theory holds that greater income inequality, by increasingly concentrating political power in the hands of the wealthy, allows them greater power to shape the political agenda in ways that discourage the broader public from taking interest (see, for example, Solt Reference Solt2008). In each of these circumstances, macrointerest is argued to increase in good, not bad, economic conditions (see also Stimson Reference Stimson2015; Peterson et al. Reference Peterson, Miller, Saunders and McClurg2022, 206).

Data to test these hypotheses are drawn from several sources. The Democratic Electoral Systems (DES) dataset updated in Bormann and Golder (Reference Bormann and Golder2022) provides information about the timing of elections, yielding a dichotomous variable coded one in election years and zero when no election was held. The three institutional variables are measured as in Kittilson and Schwindt-Bayer (Reference Kittilson and Schwindt-Bayer2010). Data on parliamentarism, a dichotomous variable coded one in pure parliamentary systems and zero otherwise, is sourced from the DES. Federalism is likewise dichotomous, coded one in countries with strong federal systems (see Lijphart Reference Lijphart1999) and zero in all others. Proportionality in the electoral system is measured using the Gallagher least-squares index of disproportionality, which measures the disparity between parties’ vote shares and their seat shares (Gallagher 1991, 40–41, Reference Gallagher2023). The context of good and bad economic conditions was measured with data on GDP per capita, national GDP growth, and unemployment from OECD.Stat (OECD 2023) and on the Gini index of disposable income inequality from the Standardized World Income Inequality Database (Solt Reference Solt2020a).

The resulting dataset comprises the thirty-seven OECD democracies, each observed in twenty-one (Mexico) to forty (Ireland, Italy, the UK, and the USA) consecutive years (mean: 33.7 years, median: 33 years). Even among these relatively data-rich countries, our measure of macrointerest provides much more data than would otherwise be available: the richest single survey for these cases, the European Social Survey, covers only 18 per cent of these country-years, does not provide annual data, and, of course, excludes entirely the nine OECD members in the Americas and around the Pacific Rim (see supplementary appendix).

Shor et al. (Reference Shor, Bafumi, Keele and Park2007) demonstrate that such pooled time series are best analyzed using a Bayesian multilevel model, including varying intercepts for each country and each year. The former help account for heteroskedasticity across space due to, for example, omitted variable bias while permitting the inclusion of time-invariant predictors such as parliamentarism and federalism. The latter take into account ‘time shocks’ that operate on all countries simultaneously (Shor et al. Reference Shor, Bafumi, Keele and Park2007, 171–72). Further, the ‘within-between random effects’ specification is employed, meaning each of the time-varying predictors is decomposed into its time-invariant country mean and the difference between each country-year value and this country mean; this specification is superior to fixed effects and other commonly used TSCS specifications for addressing omitted variable bias and endogeneity (Bell and Jones Reference Bell and Jones2015). The time-varying difference variables capture the short-term effects of the predictors, while the time-invariant country-mean variables reflect their – often different – long-run, ‘historical’ effects (Bell and Jones Reference Bell and Jones2015, 137). Moreover, as we employ a Bayesian analysis, it is straightforward to incorporate the measurement uncertainty in the data for both macrointerest and income inequality directly into the model, with the estimated values of these variables treated as random draws from distributions with unknown true means but known standard deviations (McElreath Reference McElreath2016, 425–431; see also Kurz Reference Kurz2023, 15.1.2). The model was estimated using the brms R package (Bürkner Reference Bürkner2017).

Figure 5 displays the results.Footnote ³ Consistent with the argument that campaigns bring attention-grabbing information to the public, macrointerest in election years is found to be 0.8 points (95 per cent credible interval: 0.2 to 1.4 points) higher than in years without elections. This accords with previous research finding small but well-estimated increases in political interest in election years (see, for example, Larsen Reference Larsen2022).

Figure 5. Predicting Macrointerest in OECD Democracies.

The hypothesis that power-sharing institutions yield more public interest in politics is also supported. Macrointerest is estimated to be 5.9 (95 per cent confidence interval [c.i.]: 3.6 to 8.2) points higher in countries with parliamentary systems. The point estimate for the difference in macrointerest between countries with and without federalism is estimated be 6.6 points, with 95.8 per cent of the posterior distribution greater than zero. And although disproportionality is not estimated to have long-run effects that consistently distinguish countries with more or less proportional electoral results, changes in disproportionality appear to have an immediate negative effect: a two-standard-deviation increase in the Gallagher index yields 0.7 points less macrointerest (95 per cent c.i.: −1.4 to −0.1).

Regarding the debate on whether macrointerest is invigorated or instead discouraged by bad times, the evidence of our cross-national analysis of the impact of economic conditions falls on the side of the latter. Supporting modernization theory, increases in per capita GDP have a positive short-term effect on aggregate political interest, with a two-standard-deviation increase associated with 1.3 (95 per cent c.i.: 0.2 to 2.4) points more macrointerest. The point estimate for the long-term, historical effect as evidenced by differences in mean levels across countries is found to be 4.9 points, albeit with only 90.6 per cent of the posterior distribution greater than zero. As predicted by relative power theory, the long-term effects of income inequality are strongly negative, with a two-standard-deviation difference across countries associated with 6.4 points less macrointerest (95 per cent c.i.: −12.6 to −0.6 points). Year-to-year changes in income inequality are found to make little difference – it would seem that, from one perspective, the influence of the wealthy over the political agenda does not change on such a short time scale, and from the other, that the public does not react to worsening conditions in the distribution of income with greater interest in its agents’ actions. The results with regard to growth in the national economy and with regard to unemployment similarly do not provide strong evidence of either negative or positive effects. Still, taken as a whole, this evidence indicates that at least with regard to economic conditions, it is good times, not bad ones, that yield more macrointerest.