ssoar_reference_gesis-ssoar-57735_study_program_90_search

Property	Value
?:linkContext	Despite significant methodological advancements, comparative social scientists continue to face the question of how to adequately test theoretical multilevel models empirically. Hierarchical modeling has evolved into a canonical statistical technique for regressing an individual-level variable on individual-and contextual-level predictors. There is no agreement when it comes to multilevel models where the dependent variable is analytically located on the contextual level, though. Many comparative studies 'solve' this problem through measures of central tendency -such as the average -or the distribution of the data -such as percentages. They then use these aggregates as predictors for the contextual-level dependent variable (for examples, see . This approach has been criticized on both statistical and methodological grounds. demonstrated that group mean aggregation may lead to biased estimates. argued that the aggregation procedure needs to take into account the complex theoretical relationships of independent variables at different levels of analysis. When applying simple aggregation, researchers may run the risk of drawing invalid conclusions about how individual-level predictors affect contextual-level outcomes . Given these criticisms, researchers have proposed two more advanced strategies for analyzing contextual-level outcomes in multilevel models: the multilevel SEM and the two-step approach. Since multilevel SEM and the two-step approach are seldom used in comparative survey research, the article seeks to motivate researchers to improve the validity of their inferences when analyzing contextual-level outcomes by going beyond simple aggregation. In the following section, we introduce the methodological and statistical advantages of these two alternative techniques over the group means approach. In our analysis, we illustrate these advantages in an empirical study of the effect of citizens' support for democratic values at the individual level on a contextual outcome -the persistence of democracy. We draw on data from the World Values Survey and the Quality of Government project and study 98 countries between 1946 and 2014. We compare the regression coefficients and confidence intervals of our individual-level predictor -support for democratic values -on democracies' persistence when applying the three methods. Whereas we found no significant effect of support for democratic values in the model using simple group mean aggregation, citizens' support for democratic values was a significant predictor of democracies' estimated survival rate when applying multilevel SEM and the two-step approach. In the final section we therefore conclude that comparative researchers who use simple group mean aggregation when regressing a contextual outcome on individual level predictors may run the risk of wrongly rejecting their hypothesis of interest. Testing theoretical multilevel models with contextual-level outcomes poses two challenges. From a methodological point of view, researchers need to establish close correspondence between the theoretical multilevel mechanism and its empirical measurement. From a statistical perspective, they need to choose a method that is both valid and reliable for aggregating the individual-level predictors. In the following, we discuss the methodological foundations of multilevel analysis of macrolevel social phenomena. We then proceed to introduce and compare three analytical strategies for analyzing contextual level outcomes: simple manifest group mean aggregation, latent aggregation through multilevel SEM, and the two-step approach. The results of the comparison are summarized in at the end of this chapter. According to the paradigm of structural individualism , the ultimate goal of the social sciences is to explain social phenomena on the contextual -or macro -level as a consequence of individuals' social actions on the individualor micro -level. Structural individualism distinguishes three explanatory mechanisms (see ) . Situational mechanisms (1) link the objective characteristics of the social situation to the subjective expectations and evaluations of individuals. Action-formation mechanisms (2) explain individuals' actions given their subjective definition of the situation. This is a pure micro-level explanatory step. Transformational mechanisms (3) reconstruct how individuals' actions aggregate to create a new social situation. They thereby re-link the micro level to the macro level. Studying these theoretical mechanisms empirically is not straightforward. Multilevel modeling ) is a well-established statistical tool for testing situational and action formation mechanisms, that is, explanations that link social situations to individuals' expectations, evaluations, and actual decisions . By contrast, micro-to-macro (or, more technically, level-one to level-two) explanations constitute a blind spot of conventional multilevel analysis (henceforth MLA) 1 as transformational mechanisms are more difficult to analyze empirically . The simple group means approach When studying multilevel models with contextual-level outcomes, a common approach is to aggregate all level-one variables (hereafter L1) to level-two variables (hereafter L2) by computing their group-specific arithmetic means. This manifest aggregation is followed by an L2-only regression (see ). Methodologically, this method models neither situational nor action-formation mechanisms and accounts for transformational mechanisms via (manifest) aggregation (see ). Statistically, have shown that this procedure only yields valid estimates if the L1 variance of the aggregated variables is zero. If the L1 variance is larger than zero, simple group mean aggregation yields biased estimates. In cross-national comparative survey research, this The social mechanisms of social science explanations. Source: . The social mechanisms of social science explanations. Source: is often the case because individuals are sampled from a finite population and a specific constellation of individuals is selected to measure the L2 construct . Since manifest aggregation does not control for these sampling errors, the observed group average (measured, for instance, in terms of group-specific arithmetic means) may be an unreliable measure of the unobserved true group average. In addition, the observed group average completely obscures the heterogeneity within groups. Therefore, if effects of observed group averages on L2 outcomes are of interest, estimates of both these effects and of other L2 predictors are likely to be biased when applying the simple group means approach . Multilevel SEM avoids these statistical problems by replacing manifest with latent aggregation (see ). Assume that we observe a manifest L1 variable X ij for individuals i in countries j. X ij is used to predict a manifest L2 outcome Y j along with other L2 predictors P j . Following the simple group means approach, X ij is aggregated from L1 to L2 by computing group-specific arithmetic means j X • , which are not corrected for sampling error. In a second step, j X • are used to predict Y j controlled for P j (adapted from : 2 The notation by implies group mean centering of all L1 predictors to account for a reference-group effect (in their example, this is the dependence of student academic self-concept on class-average achievement). Since our substantive application does not include a reference-group effect, we present the general notation without group mean centering. In addition, we use standard multilevel notation for the L2 residual variance. The simple group means approach By contrast, multilevel SEM regards the actual group mean on L2 as an unobserved latent variable U xj (which must not be confused with L2 residual error u oj ) that can only be estimated with error by the L1 indicators . Following the conventions of SEM, the L2 latent means of the L1 observations are therefore depicted by ovals in . While the simple group means approach treats the L2 group mean as a simple composite or index score of the L1 observations, multilevel SEM assumes the unobserved latent group means to cause the observed L1 values . Multilevel SEM proceeds in two steps: First, an L2 latent variable U xj is estimated. It is assumed to be the cause of X ij at L1. In a second step, U xj is used to predict the L2 outcome Y j along with the other L2 predictors P j : The aggregated L2 construct is a measure of the unobserved true group mean. Its reliability is a function of the relative share of the L2 variance weighted by the group-specific number of observations : As in conventional hierarchical modeling, σ² x denotes the L1 part and τ² x the L2 part of the variation of the respective indicator(s), whereas n j refers to the groupspecific number of observations. By estimating a latent L2 variable U xj as in (2), the variance of the L1 indicator is partitioned into an L1 and an L2 component. Unlike simple group mean aggregation, latent aggregation takes account of the heterogeneity within each group by partitioning the L1 variance σ² x from the L2 variance τ² x . In addition, by estimating latent group means at L2, which are assumed to cause the L1 observations in each group, the multilevel SEM approach acknowledges that the L1 scores do not perfectly map the construct at the L2 level, because of measurement error . In sum, multilevel SEM replaces manifest with latent aggregation to aggregate individual-level predictors of macro-level outcomes. Like manifest aggregation, latent aggregation per se models only the transformational but not the situational and action formation mechanism. Statistically, however, latent aggregration is superior to manifest aggregation since it corrects for sampling error (see ). As a result, its estimates are less biased, thereby permitting more valid inferences regarding the effect of multilevel predictors on contextual-level outcomes. The two-step approach also deals with the methodological and statistical issues that arise when studying multilevel models with contextual-level outcomes, albeit in a different manner. summarizes its basic idea. The two-step approach builds on standard MLA. For an L1 outcome Y ij and L1 units i nested in L2 contexts j, the standard model is given by: In equation , β 0j is the regression intercept of the outcome variable, ß 1j is the regression slope of an L1 predictor, and e ij is the residual error term. In contrast to non-nested regression analysis, both random intercepts β 0j and random slopes ß 1j can be estimated for each L2 unit j by modeling them as a function of an additional L2 predictor Z j with distinct intercepts (γ 00 and γ 10 ) and regression slopes (γ 01 and γ 11 ): Equations and introduce two additional residual error components: u 0j denotes the residual error of the outcome's L2 intercept ß 0j , and u 1j denotes the residual error of the slope ß 1j between L2 units. Standard MLA only considers the case of an L1 outcome Y ij that is predicted by L1 and L2 variables X ij and Z j , respectively. proposes an extension of the standard MLA approach to study an L2 outcome Y j : Let X 1ij be the L1 explanatory variable of primary interest. In a first step, X 1ij is regressed on all other L1 and L2 predictors X 2ij , …, X nij and Z j : In a second step, the L2 residuals u 0j of this model are used as a predictor variable in an L2 regression of the L2 outcome of interest: The effect of u 0j on the L2 outcome Y can be interpreted as the aggregated effect of the L1 variable X 1 , net of both L1 and L2 covariates X 2 , …, X n and Z. The two-step approach has both statistical and methodological advantages when studying multilevel models with contextual-level outcomes (see ). Statistically, it provides a better estimate than the group mean aggregate: u 0j is a model-based estimate of the L2 variance that is already net of the L1 variance. In addition, u 0j can be adjusted for other covariates at L1 and L2. This may save degrees of freedom and circumvent collinearity issues when using u 0j as a predictor in a subsequent L2 regression. Compared to the group means approach and the multilevel SEM approach, the crucial methodological advantage of the twostep approach is its capacity to empirically model theoretical macro-micro-macro Step 1 Step 2 Situational mechanism Transformational mechamism The two-step approach explanations in their entirety. The MLA of step 1 maps both the situational and action formation mechanism through the regression of an L1 outcome on L1 and L2 predictors. Storing the L2 residuals of this MLA then maps an underlying transformational mechanism in terms of an L1-L2 aggregation. The relative statistical performance of each method can also be compared empirically. Based on previous research, we deduce two hypotheses. First, we expect that unless the L1 variance equals zero, simple group mean aggregation yields unreliable measures of the unobserved true group means. By contrast, multilevel SEM results in reliable estimates of true group means. Consequently, when group means based on simple aggregation are used as predictors of an L2 outcome, estimates of their regression coefficients may be biased : Regression coefficients of L2 predictors that are simple group means deviate in terms of a) point estimates, b) standard errors, and c) resulting significance levels from regression coefficients of L2 predictors that have been aggregated through multilevel SEM. Second, while the statistical performance of the two-step approach is less well researched, compared multilevel SEM to another two-step approach proposed by . This approach adjusts the observed group means with weights from ANOVA formulas. This is quite similar to the decomposition of variance in an empty multilevel model. observed that approach performed slightly less well than multilevel SEM. Consequently, we expect Griffin's Residual reflects the net effect of the individual-level independent variable two-step approach to yield estimates closer to multilevel SEM than to the simple group means approach: Regression coefficients of L2 predictors that have been aggregated by the two-step approach deviate less from multilevel SEM in terms of a) point estimates, b) standard errors, and c) resulting significance levels than regression coefficients of L2 predictors that are simple group means. To illustrate the methodological and statistical issues described in the previous section, we use the persistence of democracy as a substantive example. Explanations of democratic persistence pertain either to a macro-to-micro mechanism leading from the macro level to the level of individual citizens or to a micro-to-macro mechanism leading from individual citizens to the persistence of democracy at the macro level. introduces a classic model linking macro-level causes to individuals' micro-level incentives for subverting a democratic regime. Acknowledging that democratic competition produces winners and losers, he argues that "political forces comply with present defeats because they believe that the institutional framework that organizes the democratic competition will permit them to advance their interests in the future" ). Institutions are not only crucial for inspiring the belief that there will be future possibilities to advance one's interests. The given set of political and economic institutions also has distributional consequences affecting the capacities individuals have at their disposal to advance their interests . A model of democratic persistence therefore has to take into account that -under the same set of democratic rules -members of some societal groups might deem their chances of affecting future democratic outcomes to be lower than members of other societal groups. Correspondingly, classic studies have analyzed the decisive impact of economic development on both the process of successful democratization as well as democratic persistence . A second example for the macro-to-micro mechanism underlying the persistence of democracy is the idea that an ethnically divided society poses a particular challenge to democratic persistence . In countries where several ethnic groups are politically mobilized, the question of who is to legitimately take part in the democratic game is continu-ously contested. Members of ethnic minorities often see little incentive to support ruling elites, who are -in virtue of the majority principle -likely to be members of the majority group. As a result, those out of power may choose to subvert democracy because they feel permanently excluded from democratic decisions likely to reflect only the interests of the majority. A classic example of the micro-to-macro mechanism underlying the persistence of democracy is the political culture model. seminally argued that the persistence of a political regime does not rest on its formal democratic institutions alone, but also on its political culture. Succeeding studies further specified the content of political culture and its effect on democratic persistence based on ) systems theory . According to Easton, citizens' political support refers to their supportive values and attitudes toward the political community, the political regime, and political authorities . A critical amount of political support is necessary for any kind of political system to persist. Citizens' political support increases the functionality of political systems as it allows political authorities to convert demands into outputs and permits them to implement collectively binding decisions without having to resort to force . Building on , clarifies the implications of the different dimensions of political support for democratic political regimes. Support for the political authorities is crucial for their re-or de-election; support for the political system is essential for the persistence of a given type of democracy; support for democratic values is critical for the persistence of democracy in general . Thus, citizens' support for democratic values is the key factor when studying the effect of individual-level political orientations on the persistence of democracy at the macro level. Fails and Pierce (2010) tested the systems approach of the political culture model empirically. Their analysis yielded no significant relationship between citizens' support for democratic values and their rejection of authoritarian values on the one hand and the probability of a decline of democracy on the other hand. These mechanisms can be combined into a full multilevel explanation of democratic persistence (see ). From the macro to micro explanations, we take the insight that citizens' support for democratic values is likely to be affected by context-specific economic conditions and ethnic heterogeneity. From the micro to macro explanations, we take the insight that micro-level support for democratic values crucially accounts for the persistence of democracy at the macro level. Based on the data available, we analyzed the persistence of democracy from 1946 to 2014. We derived all L2 indicators from the Quality of Government standard time series data set (QoG) , which includes data on a broad range of country-level indicators over time that we could easily merge with our L1 data. To measure our L2 outcome variabledemocratic persistence -we used the democracy index developed by the Polity IV project as included in the QoG (Marshall, Gurr, & Jaggers, 2015). Polity IV's democracy index -POLITY -reports countries' level of democracy on a scale ranging from -10 (fully autocratic) to +10 (fully democratic). 5 In line with the threshold provided on the Polity IV website , we considered countries as democracies if their POLITY score ≥ 6. As for our L2 predictors, we used the following indicators: Economic development was measured using countries' annual gross domestic product (GDP). We used the log of the OECD measure of GDP per capita. Ethnic heterogeneity was 5 POLITY is a composite score that quantifies the extent to which a country exhibits democratic and authoritarian characteristics. Polity IV coders assess countries' formal political institutions in terms of five component variables -the competitiveness of political participation (1), the openness of executive recruitment (2), the competitiveness of executive recruitment (3), the constraints on the executive (4), and the regulation of political participation (5) for each country on an annual basis. Countries are assigned weighted scores for each component. These are then added up to arrive at a democracy (DEMOC) and an autocracy score (AUTOC), both of which range from 0 to 10. The autocracy score is then subtracted from the democracy score to construct POLITY . 6 We noted an inconsistency in the definition of the thresholds. In their codebook, A two-level explanation of the persistence of democracy measured using ethno-linguistic fractionalization index (ELF), a measure of the probability that two randomly chosen individuals from a particular country are members of different ethnic groups. It ranges from 0 (perfect homogeneity) to 1 (very high fractionalization). Citizens' support for democratic values and all other L1 covariates were derived from the World Values Survey (WVS). The WVS is a cross-national survey based on representative national samples investigating worldwide sociocultural and political change. For our analyses, we used the wave 6 aggregated longitudinal file, which includes more than 340,000 observations sampled in 101 countries across all available waves from 1981 to 2014. In line with previous research, support for democratic values was operationalized in terms of respondents' reply to the following question: "I'm going to describe various types of political systems and ask what you think about each as a way of governing this country. For each one, would you say it is a very good, fairly good, fairly bad or very bad way of governing this country?". For reasons of data availability, we used respondents' rejection of an authoritarian system rather than their support for a democratic system. The answer category reads: "Having a strong leader who does not have to bother with parliament and elections" (1 = 'very good'; 2 = 'fairly good'; 3 = 'bad'; 4 = 'very bad'). For our analyses, we dichotomized this variable (0 = 'good / very good' vs. 1 = 'bad / very bad'). In accordance with previous research , we controlled for individuals' age (six categories ranging from 1 = '15-24 years' to 6 = '65 and more years'), subjective assessment of social class (five categories ranging from 1 = 'lower class' to 5 = 'upper class'), and education (eight categories ranging from 1 'inadequately completed elementary education' to 8 'university with degree/higher education'). Studying the effect of L1 and L2 predictors on an L2 outcome such as the persistence of democracy poses two methodological challenges. First, choosing a method to address the L1-L2 aggregation problem; second, analyzing persistence of democracy, which is a duration variable. We compared three different strategies for solving the L1-L2 aggregation problem. First, we aggregated support for democratic values and all other L1 covariates by computing the arithmetic means for each country year (model 1). Second, we corrected for sampling error by estimating a latent aggregation of all L1 variables on L2 using multilevel SEM (model 2). 9 Third, we applied the two-step procedure 7 The formula is: where s i is the share of group i (i = 1, …, n). proposed by by regressing support for democratic values on all other L1 and L2 predictors and then using the L2 residuals of this multilevel model as a new predictor variable. We estimated not one, but several multilevel levels that were built up stepwise: The first empty model separated the L2 residuals of support for democratic values from the L1 residuals (model A1). We then added the macro level predictors GDP and ELF (models A2-A4). Finally, we added all L1 controls (model A5). Researchers typically use stepwise model building (which we also carried out in the L2-only regressions below) to make causal claims about mediator variables partialing out significant effects of previous regressors. Apart from comparing point estimates and confidence intervals between aggregation methods for the final model, we also considered it instructive to analyze a series of stepwise models in order to assess whether different aggregation methods lead to different claims about causal mediation. In addition, we chose an adequate model for predicting democratic persistence, a duration variable. The time span of interest is the persistence of a given democracy until its breakdown. Whereas some democracies may have persisted before entering the observation window (left censoring), others may have continued to persist after the observation ended (right censoring). Within the time period of analysis, the same country may have experienced multiple democratic sequences, followed by breakdowns. In order to address these issues, we used event history modeling. We considered democratic breakdown to occur if the score of democratic regimes (nested within countries) fell below the threshold of POLITY = 6. The duration until this event was measured by the total number of years a democratic system persisted from 1946 onwards. Multiple breakdowns within the same country were coded as distinct events. To keep the models parsimonious, we used a simple exponential event history model, which assumes constant transition rates across years. In formal terms, our event history model is defined as follows: Let h denote the hazard rate of democracies' estimated risk of falling below POLITY = 6 and t the time of democracies' survival. The basic exponential survival model can then be described as: λ is a positive constant constraining transition rate (in terms of democratic breakdowns) that is equal across years. Our aim was to predict the expected survival time E(t) with an aggregate measure of citizens' support for democratic values (DVAL), countries' GDP and ELF, as well as aggregate measures of citizens' age (AGE), subjective social class (SCLASS), and education (EDUC). When applying simple aggregation, democracies' expected time of survival was estimated by: Here, U refers to the unobserved latent L2 group mean which is assumed to cause the observed L1 values of each variable. Finally, when employing the two-step approach, the estimates were derived as follows: In equation , u 0jm denotes the L2 residuals from a hierarchical regression of citizens' support for democratic values on both the L2 predictors and the L1 covariates. The subscript m indicates that the hierarchical models were built up in a stepwise manner, which is why we estimated several terms for u 0 . These formal specifications require a methodological addendum: While we estimated three L2 event history analyses after having applied each of the three aggregation methods, our theoretical explanation emphasizes the importance of citizens' support for democratic values on L1. Hence, though the event history models applied L2-only regressions, in line with the paradigm of structural individualism, we assume that the theoretical mechanisms operate via citizens' preferences and beliefs on the micro level. In line with the aim of our article, we sought to determine how the three different aggregation methods map these L1 processes when predicting an L2 outcome. In order to increase our statistical power, we used both inter-and extrapolation techniques for our independent variables. We interpolated missing values between observation points, using the -ipolate-command in Stata. In addition, we extrapolated missing values between the last valid observation and 2015, using a 'non-linear trend' scenario. We first estimated a polynomial regression of the interpolated values of each predictor on years of observations using the -lpoly-command in Stata. We then used out-of-sample predicted values to replace missing observation for subsequent years over countries. Prior to computing the comprehensive multivariate models, we compared the survival functions of democracies with high vs. low average support for democratic values. We dichotomized the support variable and compared countries with one standard deviation above vs. below the grand mean of the aggregated variable. We then compared the survival functions of these two groups of countries using group mean aggregation, the two-step model, and latent aggregation. Independent of the method of aggregation, in the long run, the estimated survival rate for democracies scoring one standard deviation above the grand mean of support for democratic values was higher than for their lower-scoring counterparts (see , appendix). Apart from a lower estimate of the survival rate of countries whose citizens had less support for democratic values in the two-step model, the differences between the aggregation methods appeared to be negligible. presents the results of the analyses using the simple group means approach (model 1), multilevel SEM (model 2), and the two-step approach (model 3). It shows both point estimates and confidence intervals for the L1 and L2 predictors. Our survival models were built up stepwise: In model 1a and 2a, the survival rate of democracies was first predicted by support for democratic values only; in model 3a, it was predicted by the L2 residuals from the multilevel null model, which separated the variance of the L1 support variable without having included any other L1 or L2 predictor. In models 1b and models 2b, we simultaneously added GDP and ELF. Correspondingly, in model 3b we included the residuals corrected for these L2 predictors. Finally, in model 1c and 2c, we added the L1 covariates; in model 3c we included the residuals corrected for the L1 covariates. Because of the low number of events, we displayed confidence intervals both on the 10% (\|t\| > 1.64; see ticks of confidence bands) and the 5% significance level (\|t\| > 1.96; see ends of confidence bands). When applying the simple group means approach, support for democratic values did not turn out to be a significant predictor of democratic survival. Point estimates varied between -3.734 in model 1a and -3.367 in model 1c, but neither 11 The overlap of valid observations for both democratic persistence and support for democratic values before and after interpolation is displayed in (appendix). The basic survivor function of democratic persistence for our reduced sample of analysis is sufficiently similar to the survivor function of the total country sample (see , appendix). As a sensitivity check, we also extrapolated our interpolated values by repeating the last valid observation of each predictor for subsequent years with missing values. Results based on this extrapolation technique are very similar to the results reported in the results section (see , appendix). Point estimates and confidence intervals of countries' democratic survival across aggregation methods. N= 917 observations, N= 122 subjects, N= 5 failures in all models estimate was larger than 1.65 times its standard error (also see , appendix). The latter also applies to all other L2 predictors and to the L1 covariates. We observed significant intercept variation in model 1a, which only included support for democratic values as a predictor variable, but not in models 1b and 1c, which controlled for the remaining variables. Values of AIC and BIC as indicators of model fit show that not much was gained by adding predictors of democratic survival apart from citizens' support for democratic values (see , appendix). When using the latent aggregation approach, the estimated confidence intervals of support for democratic values became more precise and we observed two effects of support for democratic values on democratic survival that were greater than 1.65 times their standard error (models 2a and 2b). Once the aggregated L1 covariates were controlled for, our predictor was no longer significantly associated with the outcome. Point estimates were remarkably lower after latent aggregation, ranging from -.911 in model 2a to -1.009 in model 2b (see , appendix). Having controlled for L2 structural conditions (in terms of GDP and ELF), the effect of support for democratic values became more negative from model 2a to model 2b -which points to a suppressor effect. Yet, similar to the simple group means analysis, none of the remaining variables turned out to be significant predictors of democratic survival. Model fit indices again supported the most parsimonious model 2a and intercept variation was significant in the first two submodels only. When applying the two-step approach, point estimates of support for democratic values on democratic stability were predicted with similar precision as in latent aggregation when looking at the confidence intervals. Yet, in the two-step model, we observed three significant effects at the 10% level. The L2 (u oj ) residuals of support for democratic values predicted democratic survival independent of whether they were adjusted for other L1 or L2 variables. Effect sizes ranged from -.754 in model 3a to -.651 in model 3c (see , appendix). In contrast to simple group mean and latent aggregation, the intercept remained significant in all three sub-models. Though model fit indices supported the most parsimonious model 3a, the differences between model fit indices across models were less striking than in the event history regressions following manifest and latent aggregation. Our results can be summarized as follows: In each estimation, support for democratic values was negatively associated with the event of democratic breakdown, as expected by theory. This replicated our bivariate analysis where democracies with higher support for democratic values showed a longer estimated survival rate on average. Apart from this similarity, there are notable differences between the aggregation methods: While support for democratic values was not significantly associated with democratic stability after manifest aggregation, significant effects could be observed after both latent aggregation and the two-step approach. Applying more advanced aggregation methods led to smaller point estimates and standard errors compared to the simple group means approach. All this is in line with the two hypotheses postulating notable differences between simple group means aggregation and latent aggregation, and closer similarity between the two-step approach and latent aggregation than between the two-step approach and manifest aggregation. Yet, compared to latent aggregation, which has already been observed to yield unbiased point estimates in simulation models , researchers who apply the two-step approach may run the risk of committing type one errors: In the most comprehensive model of the two-step approach (model 3c) and unlike in the corresponding regressions following latent aggregation (model 2c), the effect of support for democratic values was significant at the 10% level. In this paper, we addressed a methodological challenge well known to comparative survey researchers: how to study the effect of level two (L2) and level one (L1) predictors of a level two (L2) outcome so as to yield both reliable and valid results. Researchers have criticized simple aggregation for methodological and statistical reasons. Building on these insights and using the persistence of democracy as a substantive example, we compared the simple group means approach with two more advanced analytical strategies: the multilevel SEM approach, which estimates a latent L2 variable assumed to cause its L1 indicators, and a two-step approach, which relies on the L2 residuals of a multilevel model estimated prior to the analysis of interest . Our study corroborates previous critiques of the simple group-means approach. In both bivariate comparisons of countries' survival curves and more comprehensive multivariate event history analyses, we observed that support for democratic values was negatively associated with democratic breakdown. Unlike in the bivariate models, however, the multivariate models revealed that the associated significance levels of the estimates of support for democratic values differed remarkably depending on the aggregation method. Whereas support for democratic values was not significant in the regressions following simple group mean aggregation, confidence intervals suggested point estimates of higher precision when using either the multilevel SEM or the two-step approach, and the latter two approaches showed several significant effects at the 10% level. These empirical results show that researchers can improve the validity of their inferences by choosing more advanced analytical strategies. First, the results match previous findings from simulation analyses , which show that the simplest form of aggregation -manifest group means -is prone to beta or typetwo errors in terms of false negative findings. Second, our results challenge finding (based on simple aggregation) that support for democratic values has no effect on democracies' probability of decline. Our results suggest that comparative survey researchers interested in the effect of one or more L1 predictors on an L2 outcome may overestimate the standard errors of their regression coefficients when using manifest group mean aggregation. The two more advanced analytical strategies have distinct methodological and statistical advantages. From a statistical perspective, the two-step approach performs somewhat poorer than the multilevel SEM approach: Given that simulation revealed regression coefficients after latent aggregation to be unbiased , researchers who apply the two-step approach may run the risk of committing type-one errors in terms of false positive findings. An evident methodological advantage of the two-step approach is, however, that it is particularly suited to simultaneously model situational, action formation, and transformational mechanisms in their entirety. We conclude with several suggestions for future research. As of yet, no simulation analyses (similar to the ones comparing the simple group mean and the multilevel SEM approach) have been carried out for the two-step approach. It is therefore not possible to determine whether the estimated confidence intervals of the two-step approach are more or less reliable than the results of the latent aggregation approach. Hence, our first suggestion for future research is to perform a simulation analyses for all three aggregation methods. Controlling the data-generating mechanism would permit valid conclusions about the actual precision of each aggregation method compared to the 'real' effect size at L2. Second, the latent aggregation model can be extended towards a doubly-latent model with controls for measurement error. Thus, our second suggestion for future research is to use multiple indicators of political support to arrive at a doubly-latent model of political support at L2. Depending on the results of the aforementioned simulation study, latent variable models and the two-step approach could eventually also be combined in order to estimate both situational and transformational mechanisms without falling prey to either measurement or sampling error. Moreover, if individuals' actual decisions such as turning out to vote or participating in demonstrations or public protests are considered, a combined framework of structural equation modeling and the two-step approach would allow researchers to map action-formation mechanisms as well. 13 Third, while we used a simple exponential event-history model to simplify the analysis, future research might make use of 13 Structural equation modeling can map action formation mechanisms in simple L1 regressions as well. In addition, for group-mean centered L1 variables, multilevel SEM can estimate situational mechanisms by computing the difference between L2 and L1 regression coefficients . more flexible links for the survival function such as piecewise constant or frailty models. In sum, we encourage comparative survey researchers to surpass the simple group means aggregation approach in favor of more advanced methods of analyzing contextual-level outcomes. We have shown that this helps researchers to circumvent beta or type-two errors in terms of false negative findings when using one or more L1 indicator to predict an L2 outcome. In addition, unlike the simple group means approach, these more advanced methods can be extended further, thereby facilitating the test of more theoretically valid models. Notes. Random intercept model (QR decomposition) across country-years (level 2). Significance levels: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Distribution of democratic persistence and support for democratic values across country years A comparison of democracies' estimated survival rates across different samples of analysis Survival of democracies by support for democratic values across aggregation methods Point estimates and confidence intervals of countries' democratic survival across aggregation methods (constant interpolation) The civic culture: Political attitudes and democracy in five nations Contextualizing cognitive consonance by a social mechanisms explanation: Moderators of selective exposure in media usage Analyse & Kritik Micro-macro multilevel analysis for discrete data: A latent variable approach and an application on personal network data Sociological Methods & Research Stepwise latent class models for explaining group-level outcomes using discrete individual-level predictors Multivariate Behavioral Research Political democracy and the timing of development American Sociological Review Economic and noneconomic determinants of political democracy in the 1960s Research in political sociology Hierarchical linear models: Applications and data analysis methods Predicting group-level outcome variables from variables measured at the individual level: A latent variable multilevel model Psychological Methods Democratic challenges, democratic choices: The erosion of political support in advanced industrial democracies Index construction with formative indicators: An alternative to scale development Journal of Marketing Research A systems analysis of political life A re-assessment of the concept of political support British Journal of Political Science Changing mass attitudes and democratic deepening Political Research Quarterly Ethnic and cultural diversity by country Journal of Economic Growth The Oxford handbook of political behavior Interaction between individuals and situations: Using HLM procedures to estimate reciprocal relationships Journal of Management Social mechanisms: An analytical approach to social theory Causal mechanisms in the social sciences. Annual Review of Sociology Ethnic groups in conflict Multilevel analysis: Techniques and applications Linking societal and psychological factors to homicide rates across nations Journal of Cross-Cultural Psychology Some social requisites of democracy: Economic development and political legitimacy The American Political Science Review The multilevel latent covariate model: A new, more reliable approach to grouplevel effects in contextual studies Psychological Methods Doubly-latent models of school contextual effects: Integrating multilevel and structural equation approaches to control measurement and sampling error Multivariate Behavioral Research Polity IV Project Dataset users' manual Polity IV Project: Regime characteristics and transitions Civic culture and democracy: The question of causal relationships American Political Science Review Mplus User's Guide. Seventh Edition Critical citizens: Global support for democratic government Oxford Social expenditure: Reference series Modeling micro-macro relationships: Problems and solutions Journal of Mathematical Sociology A general multilevel SEM framework for assessing multilevel mediation Psychological Methods Democracy and the market: Political and economic reforms in Eastern Europe and Latin America Democracy and development: Political institutions and well-being in the world Politics in plural societies: A theory of democratic instability Micro-macro links and microfoundations in sociology The Journal of Mathematical Sociology Democracy in divided societies: Electoral engineering for conflict management The consolidation of democracy: Comparing Europe and Latin America A Latent Cluster-Mean Approach to the Contextual Effects Model With Missing Data Journal of Educational and Behavioral Statistics Multilevel analysis: An introduction to basic and advanced multilevel modeling The quality of government standard dataset (Version Jan16) Explaining and understanding by answering "why" and "how" questions: A programmatic introduction to the special issue "social mechanisms Analyse & Kritik The changing face of methodological individualism Annual Review of Sociology World Values Survey 1981-2014 Longitudinal Aggregate (v. 20150418) Surpassing Simple Aggregation: Advanced Strategies for Analyzing Contextual-Level Outcomes in Multilevel Models Dominik Becker dominik.becker@ifsoz.uni-tuebingen.de University of Tuebingen Wiebke Breustedt University of Duisburg Essen University of Cologne Christina Isabel Zuber University of Konstanz Institute of Sociology University of Tuebingen Wilhelmstr. 36 72074 Tuebingen Germany Surpassing Simple Aggregation: Advanced Strategies for Analyzing Contextual-Level Outcomes in Multilevel Models 10.12758/mda.2017.05 GROBID - A machine learning software for extracting information from scholarly documents transformational mechanisms contextual level outcomes multilevel analysis sampling error democratic stability democratic values This article introduces two advanced analytical strategies for analyzing contextual-level outcomes in multilevel models: the multilevel SEM and the two-step approach. Since these strategies are seldom used in comparative survey research, we first discuss their methodological and statistical advantages over the more commonly applied approach of group mean aggregation. We then illustrate these advantages in an empirical analysis of the effect of citizens' support for democratic values at the individual level on a contextual-level outcome-the persistence of democracy-drawing on data from the World Values Survey and the Quality of Government project. Whereas we found no significant effect of support for democratic values in the model using simple group mean aggregation, citizens' support for democratic values was a significant predictor of democracies' estimated survival rate when applying latent aggregation in multilevel SEM and the two-step approach. The article corroborates previous concerns with simple aggregation and demonstrates how researchers can improve the validity of their analyses of contextual-level outcomes by using alternative strategies of aggregation. Despite significant methodological advancements, comparative social scientists continue to face the question of how to adequately test theoretical multilevel models empirically. Hierarchical modeling has evolved into a canonical statistical technique for regressing an individual-level variable on individual-and contextual-level predictors. There is no agreement when it comes to multilevel models where the dependent variable is analytically located on the contextual level, though. Many comparative studies 'solve' this problem through measures of central tendency -such as the average -or the distribution of the data -such as percentages. They then use these aggregates as predictors for the contextual-level dependent variable (for examples, see Fails & Pierce, 2010; Lim, Bond, & Bond, 2005; Muller & Seligson, 1994) . This approach has been criticized on both statistical and methodological grounds. Croon and van Veldhoven (2007) demonstrated that group mean aggregation may lead to biased estimates. Griffin (1997) argued that the aggregation procedure needs to take into account the complex theoretical relationships of independent variables at different levels of analysis. When applying simple aggregation, researchers may run the risk of drawing invalid conclusions about how individual-level predictors affect contextual-level outcomes (Snijders & Bosker, 1999) . Given these criticisms, researchers have proposed two more advanced strategies for analyzing contextual-level outcomes in multilevel models: the multilevel SEM and the two-step approach. Since multilevel SEM and the two-step approach are seldom used in comparative survey research, the article seeks to motivate researchers to improve the validity of their inferences when analyzing contextual-level outcomes by going beyond simple aggregation. In the following section, we introduce the methodological and statistical advantages of these two alternative techniques over the group means approach. In our analysis, we illustrate these advantages in an empirical study of the effect of citizens' support for democratic values at the individual level on a contextual outcome -the persistence of democracy. We draw on data from the World Values Survey and the Quality of Government project and study 98 countries between 1946 and 2014. We compare the regression coefficients and confidence intervals of our individual-level predictor -support for democratic values -on democracies' persistence when applying the three methods. Whereas we found no significant effect of support for democratic values in the model using simple group mean aggregation, citizens' support for democratic values was a significant predictor of democracies' estimated survival rate when applying multilevel SEM and the two-step approach. In the final section we therefore conclude that comparative researchers who use simple group mean aggregation when regressing a contextual outcome on individual level predictors may run the risk of wrongly rejecting their hypothesis of interest. Methodological Foundation and Statistical Background Testing theoretical multilevel models with contextual-level outcomes poses two challenges. From a methodological point of view, researchers need to establish close correspondence between the theoretical multilevel mechanism and its empirical measurement. From a statistical perspective, they need to choose a method that is both valid and reliable for aggregating the individual-level predictors. In the following, we discuss the methodological foundations of multilevel analysis of macrolevel social phenomena. We then proceed to introduce and compare three analytical strategies for analyzing contextual level outcomes: simple manifest group mean aggregation, latent aggregation through multilevel SEM, and the two-step approach. The results of the comparison are summarized in Table 1 at the end of this chapter. Methodological Foundation According to the paradigm of structural individualism (Udehn, 2002) , the ultimate goal of the social sciences is to explain social phenomena on the contextual -or macro -level as a consequence of individuals' social actions on the individualor micro -level. Structural individualism distinguishes three explanatory mechanisms (see Figure 1 ) (Hedström & Swedberg, 1998; Tranow, Beckers, & Becker, 2016) . Situational mechanisms (1) link the objective characteristics of the social situation to the subjective expectations and evaluations of individuals. Action-formation mechanisms (2) explain individuals' actions given their subjective definition of the situation. This is a pure micro-level explanatory step. Transformational mechanisms (3) reconstruct how individuals' actions aggregate to create a new social situation. They thereby re-link the micro level to the macro level. Studying these theoretical mechanisms empirically is not straightforward. Multilevel modeling (Bryk & Raudenbush, 1992; Hox, 2010 ) is a well-established statistical tool for testing situational and action formation mechanisms, that is, explanations that link social situations to individuals' expectations, evaluations, and actual decisions (Becker, Beckers, Franzmann, & Hagenah, 2016) . By contrast, micro-to-macro (or, more technically, level-one to level-two) explanations constitute a blind spot of conventional multilevel analysis (henceforth MLA) 1 as transformational mechanisms are more difficult to analyze empirically (Opp, 2011; Raub, Buskens, & van Assen, 2011) . Three Analytical Strategies The simple group means approach When studying multilevel models with contextual-level outcomes, a common approach (Lim et al., 2005) is to aggregate all level-one variables (hereafter L1) to level-two variables (hereafter L2) by computing their group-specific arithmetic means. This manifest aggregation is followed by an L2-only regression (see Figure 2 ). Methodologically, this method models neither situational nor action-formation mechanisms and accounts for transformational mechanisms via (manifest) aggregation (see Figure 2 ). Statistically, Croon and van Veldhoven (2007) have shown that this procedure only yields valid estimates if the L1 variance of the aggregated variables is zero. If the L1 variance is larger than zero, simple group mean aggregation yields biased estimates. In cross-national comparative survey research, this Figure 1 The social mechanisms of social science explanations. Source: Hedström & Ylikoski (2010, p. 53) . Figure 1 The social mechanisms of social science explanations. Source: Hedström & Ylikoski (2010, p. 53) is often the case because individuals are sampled from a finite population and a specific constellation of individuals is selected to measure the L2 construct (Lüdtke et al., 2008) . Since manifest aggregation does not control for these sampling errors, the observed group average (measured, for instance, in terms of group-specific arithmetic means) may be an unreliable measure of the unobserved true group average. In addition, the observed group average completely obscures the heterogeneity within groups. Therefore, if effects of observed group averages on L2 outcomes are of interest, estimates of both these effects and of other L2 predictors are likely to be biased when applying the simple group means approach (Bennink et al., 2013 (Bennink et al., , 2015 Shin & Raudenbush, 2010) . The multilevel SEM approach Multilevel SEM avoids these statistical problems by replacing manifest with latent aggregation (see Figure 3 ). Assume that we observe a manifest L1 variable X ij for individuals i in countries j. X ij is used to predict a manifest L2 outcome Y j along with other L2 predictors P j . Following the simple group means approach, X ij is aggregated from L1 to L2 by computing group-specific arithmetic means j X • , which are not corrected for sampling error. In a second step, j X • are used to predict Y j controlled for P j (adapted from Marsh et al., 2009) : 2 0 1 • 2 0 j j j j Y X P u β β β = + + +(1) 2 The notation by Marsh et al. (2009) implies group mean centering of all L1 predictors to account for a reference-group effect (in their example, this is the dependence of student academic self-concept on class-average achievement). Since our substantive application does not include a reference-group effect, we present the general notation without group mean centering. In addition, we use standard multilevel notation for the L2 residual variance. Figure 2 The simple group means approach By contrast, multilevel SEM regards the actual group mean on L2 as an unobserved latent variable U xj (which must not be confused with L2 residual error u oj ) that can only be estimated with error by the L1 indicators (Marsh et al., 2009) . Following the conventions of SEM, the L2 latent means of the L1 observations are therefore depicted by ovals in Figure 3 . While the simple group means approach treats the L2 group mean as a simple composite or index score of the L1 observations, multilevel SEM assumes the unobserved latent group means to cause the observed L1 values (Lüdtke et al., 2008) . 3 Multilevel SEM proceeds in two steps: First, an L2 latent variable U xj is estimated. It is assumed to be the cause of X ij at L1. In a second step, U xj is used to predict the L2 outcome Y j along with the other L2 predictors P j : 4 0 1 2 0 j xj j j Y U P u β β β = + + +(2) The aggregated L2 construct is a measure of the unobserved true group mean. Its reliability is a function of the relative share of the L2 variance weighted by the group-specific number of observations (Lüdtke et al., 2008) : As in conventional hierarchical modeling, σ² x denotes the L1 part and τ² x the L2 part of the variation of the respective indicator(s), whereas n j refers to the groupspecific number of observations. By estimating a latent L2 variable U xj as in (2), the variance of the L1 indicator is partitioned into an L1 and an L2 component. Unlike simple group mean aggregation, latent aggregation takes account of the heterogeneity within each group by partitioning the L1 variance σ² x from the L2 variance τ² x . In addition, by estimating latent group means at L2, which are assumed to cause the L1 observations in each group, the multilevel SEM approach acknowledges that the L1 scores do not perfectly map the construct at the L2 level, because of measurement error (Bennink et al., 2013 (Bennink et al., , 2015 Preacher et al., 2010) . ( ) 2 2 2 / x x x j n τ τ σ +(3) In sum, multilevel SEM replaces manifest with latent aggregation to aggregate individual-level predictors of macro-level outcomes. Like manifest aggregation, latent aggregation per se models only the transformational but not the situational and action formation mechanism. Statistically, however, latent aggregration is superior to manifest aggregation since it corrects for sampling error (see Table 1 ). As a result, its estimates are less biased, thereby permitting more valid inferences regarding the effect of multilevel predictors on contextual-level outcomes. The two-step approach The two-step approach also deals with the methodological and statistical issues that arise when studying multilevel models with contextual-level outcomes, albeit in a different manner. Figure 4 summarizes its basic idea. The two-step approach builds on standard MLA. For an L1 outcome Y ij and L1 units i nested in L2 contexts j, the standard model is given by: 0 1 ij j j ij ij Y X e β β = + +(4) In equation 4 , β 0j is the regression intercept of the outcome variable, ß 1j is the regression slope of an L1 predictor, and e ij is the residual error term. In contrast to non-nested regression analysis, both random intercepts β 0j and random slopes ß 1j can be estimated for each L2 unit j by modeling them as a function of an additional L2 predictor Z j with distinct intercepts (γ 00 and γ 10 ) and regression slopes (γ 01 and γ 11 ): 0 00 01 0 j j j Z u β γ γ = + + (5) 1 10 11 1 j j j Z u β γ γ = + +(6) Equations 5 and 6 introduce two additional residual error components: u 0j denotes the residual error of the outcome's L2 intercept ß 0j , and u 1j denotes the residual error of the slope ß 1j between L2 units. Standard MLA only considers the case of an L1 outcome Y ij that is predicted by L1 and L2 variables X ij and Z j , respectively. Griffin (1997) proposes an extension of the standard MLA approach to study an L2 outcome Y j : Let X 1ij be the L1 explanatory variable of primary interest. In a first step, X 1ij is regressed on all other L1 and L2 predictors X 2ij , …, X nij and Z j : 1 00 01 0 1 2 ... ij j j j ij nj nij ij X Z X X e γ γ τ β β = + + + + + + (7) In a second step, the L2 residuals u 0j of this model are used as a predictor variable in an L2 regression of the L2 outcome of interest: 0 1 0 j j j Y u e β β = + + (8) The effect of u 0j on the L2 outcome Y can be interpreted as the aggregated effect of the L1 variable X 1 , net of both L1 and L2 covariates X 2 , …, X n and Z. The two-step approach has both statistical and methodological advantages when studying multilevel models with contextual-level outcomes (see Table 1 ). Statistically, it provides a better estimate than the group mean aggregate: u 0j is a model-based estimate of the L2 variance that is already net of the L1 variance. In addition, u 0j can be adjusted for other covariates at L1 and L2. This may save degrees of freedom and circumvent collinearity issues when using u 0j as a predictor in a subsequent L2 regression. Compared to the group means approach and the multilevel SEM approach, the crucial methodological advantage of the twostep approach is its capacity to empirically model theoretical macro-micro-macro L1 predictor X ij L2 predictors P j Level 1 Level 2 L2 residuals (τ 0j ) L2 outcome Y j Step 1 Step 2 Situational mechanism Transformational mechamism Figure 4 The two-step approach explanations in their entirety. The MLA of step 1 maps both the situational and action formation mechanism through the regression of an L1 outcome on L1 and L2 predictors. Storing the L2 residuals of this MLA then maps an underlying transformational mechanism in terms of an L1-L2 aggregation. The relative statistical performance of each method can also be compared empirically. Based on previous research, we deduce two hypotheses. First, we expect that unless the L1 variance equals zero, simple group mean aggregation yields unreliable measures of the unobserved true group means. By contrast, multilevel SEM results in reliable estimates of true group means. Consequently, when group means based on simple aggregation are used as predictors of an L2 outcome, estimates of their regression coefficients may be biased (Bennink et al., 2013 (Bennink et al., , 2015 : H 1 : Regression coefficients of L2 predictors that are simple group means deviate in terms of a) point estimates, b) standard errors, and c) resulting significance levels from regression coefficients of L2 predictors that have been aggregated through multilevel SEM. Second, while the statistical performance of the two-step approach (Griffin, 1997) is less well researched, Lüdtke et al. (2008) compared multilevel SEM to another two-step approach proposed by Croon and van Veldhoven (2007) . This approach adjusts the observed group means with weights from ANOVA formulas. This is quite similar to the decomposition of variance in an empty multilevel model. Lüdtke et al. (2008) observed that Croon and van Veldhoven's (2007) approach performed slightly less well than multilevel SEM. Consequently, we expect Griffin's Residual reflects the net effect of the individual-level independent variable two-step approach to yield estimates closer to multilevel SEM than to the simple group means approach: H 2 : Regression coefficients of L2 predictors that have been aggregated by the two-step approach deviate less from multilevel SEM in terms of a) point estimates, b) standard errors, and c) resulting significance levels than regression coefficients of L2 predictors that are simple group means. Substantive Application: A Multilevel Explanation of the Persistence of Democracy Theoretical Background To illustrate the methodological and statistical issues described in the previous section, we use the persistence of democracy as a substantive example. Explanations of democratic persistence pertain either to a macro-to-micro mechanism leading from the macro level to the level of individual citizens or to a micro-to-macro mechanism leading from individual citizens to the persistence of democracy at the macro level. Przeworski (1991) introduces a classic model linking macro-level causes to individuals' micro-level incentives for subverting a democratic regime. Acknowledging that democratic competition produces winners and losers, he argues that "political forces comply with present defeats because they believe that the institutional framework that organizes the democratic competition will permit them to advance their interests in the future" (Przeworski, 1991, p. 19 ). Institutions are not only crucial for inspiring the belief that there will be future possibilities to advance one's interests. The given set of political and economic institutions also has distributional consequences affecting the capacities individuals have at their disposal to advance their interests (Przeworski, 1991) . A model of democratic persistence therefore has to take into account that -under the same set of democratic rules -members of some societal groups might deem their chances of affecting future democratic outcomes to be lower than members of other societal groups. Correspondingly, classic studies have analyzed the decisive impact of economic development on both the process of successful democratization (Bollen, 1979; Bollen & Jackman, 1985; Lipset, 1959) as well as democratic persistence (Przeworski, Alvarez, Cheibub, & Limongi, 2000) . A second example for the macro-to-micro mechanism underlying the persistence of democracy is the idea that an ethnically divided society poses a particular challenge to democratic persistence (Horowitz, 1985; Rabushka & Shepsle, 1972; Reilly, 2001) . In countries where several ethnic groups are politically mobilized, the question of who is to legitimately take part in the democratic game is continu-ously contested. Members of ethnic minorities often see little incentive to support ruling elites, who are -in virtue of the majority principle -likely to be members of the majority group. As a result, those out of power may choose to subvert democracy because they feel permanently excluded from democratic decisions likely to reflect only the interests of the majority. A classic example of the micro-to-macro mechanism underlying the persistence of democracy is the political culture model. Almond and Verba (1963) seminally argued that the persistence of a political regime does not rest on its formal democratic institutions alone, but also on its political culture. Succeeding studies further specified the content of political culture and its effect on democratic persistence based on Easton's (1965 Easton's ( , 1975 ) systems theory (Dalton, 2004; Fuchs, 2007; Norris, 1999) . According to Easton, citizens' political support refers to their supportive values and attitudes toward the political community, the political regime, and political authorities (Easton, 1965) . A critical amount of political support is necessary for any kind of political system to persist. Citizens' political support increases the functionality of political systems as it allows political authorities to convert demands into outputs and permits them to implement collectively binding decisions without having to resort to force (Easton, 1965) . Building on Easton (1965 Easton ( , 1975 , Fuchs (2007) clarifies the implications of the different dimensions of political support for democratic political regimes. Support for the political authorities is crucial for their re-or de-election; support for the political system is essential for the persistence of a given type of democracy; support for democratic values is critical for the persistence of democracy in general (Fuchs, 2007) . Thus, citizens' support for democratic values is the key factor when studying the effect of individual-level political orientations on the persistence of democracy at the macro level. Fails and Pierce (2010) tested the systems approach of the political culture model empirically. Their analysis yielded no significant relationship between citizens' support for democratic values and their rejection of authoritarian values on the one hand and the probability of a decline of democracy on the other hand. These mechanisms can be combined into a full multilevel explanation of democratic persistence (see Figure 5 ). From the macro to micro explanations, we take the insight that citizens' support for democratic values is likely to be affected by context-specific economic conditions and ethnic heterogeneity. From the micro to macro explanations, we take the insight that micro-level support for democratic values crucially accounts for the persistence of democracy at the macro level. Research Design Period of analysis and data Based on the data available, we analyzed the persistence of democracy from 1946 to 2014. We derived all L2 indicators from the Quality of Government standard time series data set (QoG) (Teorell et al., 2016) , which includes data on a broad range of country-level indicators over time that we could easily merge with our L1 data. To measure our L2 outcome variabledemocratic persistence -we used the democracy index developed by the Polity IV project as included in the QoG (Marshall, Gurr, & Jaggers, 2015). Polity IV's democracy index -POLITY -reports countries' level of democracy on a scale ranging from -10 (fully autocratic) to +10 (fully democratic). 5 In line with the threshold provided on the Polity IV website (Marshall & Gurr, 2014) , we considered countries as democracies if their POLITY score ≥ 6. 6 As for our L2 predictors, we used the following indicators: Economic development was measured using countries' annual gross domestic product (GDP). We used the log of the OECD measure of GDP per capita. Ethnic heterogeneity was 5 POLITY is a composite score that quantifies the extent to which a country exhibits democratic and authoritarian characteristics. Polity IV coders assess countries' formal political institutions in terms of five component variables -the competitiveness of political participation (1), the openness of executive recruitment (2), the competitiveness of executive recruitment (3), the constraints on the executive (4), and the regulation of political participation (5) for each country on an annual basis. Countries are assigned weighted scores for each component. These are then added up to arrive at a democracy (DEMOC) and an autocracy score (AUTOC), both of which range from 0 to 10. The autocracy score is then subtracted from the democracy score to construct POLITY (Marshall et al., 2015) . 6 We noted an inconsistency in the definition of the thresholds. In their codebook, Marshall et al. (2015) Figure 5 A two-level explanation of the persistence of democracy measured using Fearon's (2003) ethno-linguistic fractionalization index (ELF), a measure of the probability that two randomly chosen individuals from a particular country are members of different ethnic groups. It ranges from 0 (perfect homogeneity) to 1 (very high fractionalization). 7 Citizens' support for democratic values and all other L1 covariates were derived from the World Values Survey (WVS). The WVS is a cross-national survey based on representative national samples investigating worldwide sociocultural and political change. For our analyses, we used the wave 6 aggregated longitudinal file, which includes more than 340,000 observations sampled in 101 countries across all available waves from 1981 to 2014. In line with previous research, support for democratic values was operationalized in terms of respondents' reply to the following question: "I'm going to describe various types of political systems and ask what you think about each as a way of governing this country. For each one, would you say it is a very good, fairly good, fairly bad or very bad way of governing this country?". For reasons of data availability, we used respondents' rejection of an authoritarian system rather than their support for a democratic system. The answer category reads: "Having a strong leader who does not have to bother with parliament and elections" (1 = 'very good'; 2 = 'fairly good'; 3 = 'bad'; 4 = 'very bad'). For our analyses, we dichotomized this variable (0 = 'good / very good' vs. 1 = 'bad / very bad'). In accordance with previous research (Schneider, 2009) , we controlled for individuals' age (six categories ranging from 1 = '15-24 years' to 6 = '65 and more years'), subjective assessment of social class (five categories ranging from 1 = 'lower class' to 5 = 'upper class'), and education (eight categories ranging from 1 'inadequately completed elementary education' to 8 'university with degree/higher education'). 8 Methods of analysis Studying the effect of L1 and L2 predictors on an L2 outcome such as the persistence of democracy poses two methodological challenges. First, choosing a method to address the L1-L2 aggregation problem; second, analyzing persistence of democracy, which is a duration variable. We compared three different strategies for solving the L1-L2 aggregation problem. First, we aggregated support for democratic values and all other L1 covariates by computing the arithmetic means for each country year (model 1). Second, we corrected for sampling error by estimating a latent aggregation of all L1 variables on L2 using multilevel SEM (model 2). 9 Third, we applied the two-step procedure 7 The formula is: 2 1 1- n i i s = ∑ where s i is the share of group i (i = 1, …, n). proposed by Griffin (1997) by regressing support for democratic values on all other L1 and L2 predictors and then using the L2 residuals of this multilevel model as a new predictor variable. We estimated not one, but several multilevel levels that were built up stepwise: The first empty model separated the L2 residuals of support for democratic values from the L1 residuals (model A1). We then added the macro level predictors GDP and ELF (models A2-A4). Finally, we added all L1 controls (model A5). 10 Researchers typically use stepwise model building (which we also carried out in the L2-only regressions below) to make causal claims about mediator variables partialing out significant effects of previous regressors. Apart from comparing point estimates and confidence intervals between aggregation methods for the final model, we also considered it instructive to analyze a series of stepwise models in order to assess whether different aggregation methods lead to different claims about causal mediation. In addition, we chose an adequate model for predicting democratic persistence, a duration variable. The time span of interest is the persistence of a given democracy until its breakdown. Whereas some democracies may have persisted before entering the observation window (left censoring), others may have continued to persist after the observation ended (right censoring). Within the time period of analysis, the same country may have experienced multiple democratic sequences, followed by breakdowns. In order to address these issues, we used event history modeling. We considered democratic breakdown to occur if the score of democratic regimes (nested within countries) fell below the threshold of POLITY = 6. The duration until this event was measured by the total number of years a democratic system persisted from 1946 onwards. Multiple breakdowns within the same country were coded as distinct events. To keep the models parsimonious, we used a simple exponential event history model, which assumes constant transition rates across years. In formal terms, our event history model is defined as follows: Let h denote the hazard rate of democracies' estimated risk of falling below POLITY = 6 and t the time of democracies' survival. The basic exponential survival model can then be described as: ( ) ; 0, 0 h t t λ λ = > > (9) λ is a positive constant constraining transition rate (in terms of democratic breakdowns) that is equal across years. Our aim was to predict the expected survival time E(t) with an aggregate measure of citizens' support for democratic values (DVAL), countries' GDP and ELF, as well as aggregate measures of citizens' age (AGE), subjective social class (SCLASS), and education (EDUC). When applying simple aggregation, democracies' expected time of survival was estimated by: ( ) Here, U refers to the unobserved latent L2 group mean which is assumed to cause the observed L1 values of each variable. Finally, when employing the two-step approach, the estimates were derived as follows: ( ) ( ) 0 1 0 exp j jm E t u β β = +(12) In equation 12 , u 0jm denotes the L2 residuals from a hierarchical regression of citizens' support for democratic values on both the L2 predictors and the L1 covariates. The subscript m indicates that the hierarchical models were built up in a stepwise manner, which is why we estimated several terms for u 0 . These formal specifications require a methodological addendum: While we estimated three L2 event history analyses after having applied each of the three aggregation methods, our theoretical explanation emphasizes the importance of citizens' support for democratic values on L1. Hence, though the event history models applied L2-only regressions, in line with the paradigm of structural individualism, we assume that the theoretical mechanisms operate via citizens' preferences and beliefs on the micro level. In line with the aim of our article, we sought to determine how the three different aggregation methods map these L1 processes when predicting an L2 outcome. In order to increase our statistical power, we used both inter-and extrapolation techniques for our independent variables. We interpolated missing values between observation points, using the -ipolate-command in Stata. In addition, we extrapolated missing values between the last valid observation and 2015, using a 'non-linear trend' scenario. We first estimated a polynomial regression of the interpolated values of each predictor on years of observations using the -lpoly-command in Stata. We then used out-of-sample predicted values to replace missing observation for subsequent years over countries. 11 Results Prior to computing the comprehensive multivariate models, we compared the survival functions of democracies with high vs. low average support for democratic values. We dichotomized the support variable and compared countries with one standard deviation above vs. below the grand mean of the aggregated variable. We then compared the survival functions of these two groups of countries using group mean aggregation, the two-step model, and latent aggregation. Independent of the method of aggregation, in the long run, the estimated survival rate for democracies scoring one standard deviation above the grand mean of support for democratic values was higher than for their lower-scoring counterparts (see Figure A3 , appendix). Apart from a lower estimate of the survival rate of countries whose citizens had less support for democratic values in the two-step model, the differences between the aggregation methods appeared to be negligible. Figure 6 presents the results of the analyses using the simple group means approach (model 1), multilevel SEM (model 2), and the two-step approach (model 3). It shows both point estimates and confidence intervals for the L1 and L2 predictors. Our survival models were built up stepwise: In model 1a and 2a, the survival rate of democracies was first predicted by support for democratic values only; in model 3a, it was predicted by the L2 residuals from the multilevel null model, which separated the variance of the L1 support variable without having included any other L1 or L2 predictor. In models 1b and models 2b, we simultaneously added GDP and ELF. Correspondingly, in model 3b we included the residuals corrected for these L2 predictors. Finally, in model 1c and 2c, we added the L1 covariates; in model 3c we included the residuals corrected for the L1 covariates. Because of the low number of events, we displayed confidence intervals both on the 10% (\|t\| > 1.64; see ticks of confidence bands) and the 5% significance level (\|t\| > 1.96; see ends of confidence bands). When applying the simple group means approach, support for democratic values did not turn out to be a significant predictor of democratic survival. Point estimates varied between -3.734 in model 1a and -3.367 in model 1c, but neither 11 The overlap of valid observations for both democratic persistence and support for democratic values before and after interpolation is displayed in Figure A1 (appendix). The basic survivor function of democratic persistence for our reduced sample of analysis is sufficiently similar to the survivor function of the total country sample (see Figure A2 , appendix). As a sensitivity check, we also extrapolated our interpolated values by repeating the last valid observation of each predictor for subsequent years with missing values. Results based on this extrapolation technique are very similar to the results reported in the results section (see Figure A4 , appendix). Figure 6 Point estimates and confidence intervals of countries' democratic survival across aggregation methods. N= 917 observations, N= 122 subjects, N= 5 failures in all models estimate was larger than 1.65 times its standard error (also see Table A3 , appendix). The latter also applies to all other L2 predictors and to the L1 covariates. We observed significant intercept variation in model 1a, which only included support for democratic values as a predictor variable, but not in models 1b and 1c, which controlled for the remaining variables. Values of AIC and BIC as indicators of model fit show that not much was gained by adding predictors of democratic survival apart from citizens' support for democratic values (see Table A3 , appendix). When using the latent aggregation approach, the estimated confidence intervals of support for democratic values became more precise and we observed two effects of support for democratic values on democratic survival that were greater than 1.65 times their standard error (models 2a and 2b). Once the aggregated L1 covariates were controlled for, our predictor was no longer significantly associated with the outcome. Point estimates were remarkably lower after latent aggregation, ranging from -.911 in model 2a to -1.009 in model 2b (see Table A4 , appendix). Having controlled for L2 structural conditions (in terms of GDP and ELF), the effect of support for democratic values became more negative from model 2a to model 2b -which points to a suppressor effect. Yet, similar to the simple group means analysis, none of the remaining variables turned out to be significant predictors of democratic survival. Model fit indices again supported the most parsimonious model 2a and intercept variation was significant in the first two submodels only. When applying the two-step approach, point estimates of support for democratic values on democratic stability were predicted with similar precision as in latent aggregation when looking at the confidence intervals. Yet, in the two-step model, we observed three significant effects at the 10% level. The L2 (u oj ) residuals of support for democratic values predicted democratic survival independent of whether they were adjusted for other L1 or L2 variables. Effect sizes ranged from -.754 in model 3a to -.651 in model 3c (see Table A5 , appendix). In contrast to simple group mean and latent aggregation, the intercept remained significant in all three sub-models. Though model fit indices supported the most parsimonious model 3a, the differences between model fit indices across models were less striking than in the event history regressions following manifest and latent aggregation. Our results can be summarized as follows: In each estimation, support for democratic values was negatively associated with the event of democratic breakdown, as expected by theory. This replicated our bivariate analysis where democracies with higher support for democratic values showed a longer estimated survival rate on average. Apart from this similarity, there are notable differences between the aggregation methods: While support for democratic values was not significantly associated with democratic stability after manifest aggregation, significant effects could be observed after both latent aggregation and the two-step approach. Applying more advanced aggregation methods led to smaller point estimates and standard errors compared to the simple group means approach. All this is in line with the two hypotheses postulating notable differences between simple group means aggregation and latent aggregation, and closer similarity between the two-step approach and latent aggregation than between the two-step approach and manifest aggregation. Yet, compared to latent aggregation, which has already been observed to yield unbiased point estimates in simulation models (Bennink et al., 2013 (Bennink et al., , 2015 Lüdtke et al., 2008) , researchers who apply the two-step approach may run the risk of committing type one errors: In the most comprehensive model of the two-step approach (model 3c) and unlike in the corresponding regressions following latent aggregation (model 2c), the effect of support for democratic values was significant at the 10% level. 12 Conclusion In this paper, we addressed a methodological challenge well known to comparative survey researchers: how to study the effect of level two (L2) and level one (L1) predictors of a level two (L2) outcome so as to yield both reliable and valid results. Researchers have criticized simple aggregation for methodological and statistical reasons. Building on these insights and using the persistence of democracy as a substantive example, we compared the simple group means approach with two more advanced analytical strategies: the multilevel SEM approach, which estimates a latent L2 variable assumed to cause its L1 indicators, and a two-step approach, which relies on the L2 residuals of a multilevel model estimated prior to the analysis of interest (Griffin, 1997) . Our study corroborates previous critiques of the simple group-means approach. In both bivariate comparisons of countries' survival curves and more comprehensive multivariate event history analyses, we observed that support for democratic values was negatively associated with democratic breakdown. Unlike in the bivariate models, however, the multivariate models revealed that the associated significance levels of the estimates of support for democratic values differed remarkably depending on the aggregation method. Whereas support for democratic values was not significant in the regressions following simple group mean aggregation, confidence intervals suggested point estimates of higher precision when using either the multilevel SEM or the two-step approach, and the latter two approaches showed several significant effects at the 10% level. These empirical results show that researchers can improve the validity of their inferences by choosing more advanced analytical strategies. First, the results match previous findings from simulation analyses (Lüdtke et al., 2008) , which show that the simplest form of aggregation -manifest group means -is prone to beta or typetwo errors in terms of false negative findings. Second, our results challenge Fails and Pierce's (2010) finding (based on simple aggregation) that support for democratic values has no effect on democracies' probability of decline. Our results suggest that comparative survey researchers interested in the effect of one or more L1 predictors on an L2 outcome may overestimate the standard errors of their regression coefficients when using manifest group mean aggregation. The two more advanced analytical strategies have distinct methodological and statistical advantages. From a statistical perspective, the two-step approach performs somewhat poorer than the multilevel SEM approach: Given that simulation revealed regression coefficients after latent aggregation to be unbiased (Bennink et al., 2013 (Bennink et al., , 2015 Lüdtke et al., 2008) , researchers who apply the two-step approach may run the risk of committing type-one errors in terms of false positive findings. An evident methodological advantage of the two-step approach is, however, that it is particularly suited to simultaneously model situational, action formation, and transformational mechanisms in their entirety. We conclude with several suggestions for future research. As of yet, no simulation analyses (similar to the ones comparing the simple group mean and the multilevel SEM approach) have been carried out for the two-step approach. It is therefore not possible to determine whether the estimated confidence intervals of the two-step approach are more or less reliable than the results of the latent aggregation approach. Hence, our first suggestion for future research is to perform a simulation analyses for all three aggregation methods. Controlling the data-generating mechanism would permit valid conclusions about the actual precision of each aggregation method compared to the 'real' effect size at L2. Second, the latent aggregation model can be extended towards a doubly-latent model with controls for measurement error. Thus, our second suggestion for future research is to use multiple indicators of political support to arrive at a doubly-latent model of political support at L2. Depending on the results of the aforementioned simulation study, latent variable models and the two-step approach could eventually also be combined in order to estimate both situational and transformational mechanisms without falling prey to either measurement or sampling error. Moreover, if individuals' actual decisions such as turning out to vote or participating in demonstrations or public protests are considered, a combined framework of structural equation modeling and the two-step approach would allow researchers to map action-formation mechanisms as well. 13 Third, while we used a simple exponential event-history model to simplify the analysis, future research might make use of 13 Structural equation modeling can map action formation mechanisms in simple L1 regressions as well. In addition, for group-mean centered L1 variables, multilevel SEM can estimate situational mechanisms by computing the difference between L2 and L1 regression coefficients (Marsh et al., 2009) . more flexible links for the survival function such as piecewise constant or frailty models. In sum, we encourage comparative survey researchers to surpass the simple group means aggregation approach in favor of more advanced methods of analyzing contextual-level outcomes. We have shown that this helps researchers to circumvent beta or type-two errors in terms of false negative findings when using one or more L1 indicator to predict an L2 outcome. In addition, unlike the simple group means approach, these more advanced methods can be extended further, thereby facilitating the test of more theoretically valid models. Notes. Random intercept model (QR decomposition) across country-years (level 2). Significance levels: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Notes. Significance levels: + < .10: * < .05; < .01; * < .001 (two-sided). Standard errors in parentheses. Figure A1 Distribution of democratic persistence and support for democratic values across country years Figure A2 A comparison of democracies' estimated survival rates across different samples of analysis Figure A3 Survival of democracies by support for democratic values across aggregation methods Figure A4 Point estimates and confidence intervals of countries' democratic survival across aggregation methods (constant interpolation) Figure 3 3 Latent aggregation in multilevel structural equation modeling Table 1 1 Comparison of methods for analyzing macro-micro-macro models Main methodological Main statistical advantages & disadvantages advantages & disadvantages group mean Transformational mechanism (via Simple to perform, but only valid if aggregation manifest aggregation and macro variance of L1 variable = 0 regression) ML SEM Transformational mechanism (via Takes sampling error into account: latent aggregation and macro reduction of estimator bias regression) 2-Step 1st step: situational & action-forma- tion mechanism (via MLA) 2nd step: transformational mecha- nism (via residuals and macro regression) Table A2 A2 Multilevel logistic regression of support for democratic values (dichotomized) on level-two predictors and level-one covariates Null model Model 1a Model 1b b se b se b se Intercept 1.812 (0.585) 2.042* (0.580) 0.457* (0.071) log(GDP) -0.174 (0.061) -0.166 (0.061) ELF -0.392 (0.345) -0.427 (0.341) Age: 15-24 years REFERENCE CATEGORY 25-34 years 0.015 (0.015) 35-44 years 0.067* (0.015) 45-54 years 0.103* (0.017) 55-64 0.092* (0.019) 65 and more years -0.039 (0.020) Education: Inadequately completed elementary REFERENCE CATEGORY Completed elementary 0.042 (0.022) Incomplete secondary: tech./voc. 0.051* (0.025) Completed secondary: tech./voc. 0.178* (0.022) Incomplete secondary: univ. prep. 0.171* (0.024) Complete secondary: univ. prep. 0.274* (0.022) Some university without degree 0.428* (0.026) University with degree 0.581*** (0.023) Subjective class: lower REFERENCE CATEGORY working 0.016 (0.017) lower middle 0.042* (0.017) upper middle -0.034 (0.019) upper -0.275*** (0.038) τ 0j 0.025 (0.063) 0.012 (0.063) -0.045 (0.054) N 219740 219740 219740 AIC 261954 263445 263440 Table A3 A3 Exponential event-history regression of democratic breakdown on aggregated support for democratic values, L2 predictors, and aggregated L1 controls (simple group-means approach) Model 2a Model 2b Model 2c b/se b/se b/se Intercept -3.220* -3.662 -2.073 (1.252) (3.492) (6.503) Support for democratic values -3.734 -3.642 -3.367 (2.485) (2.754) (2.783) log(GDP) 0.01 -0.038 (0.399) (0.432) ELF 0.715 1.294 (2.131) (2.495) Age 0.662 (1.419) Education -0.315 (0.685) Subjective class -0.846 (1.887) AIC 43.318 47.201 52.375 BIC 52.96 66.486 86.123 N (failures) 5 5 5 N (subjects) 122 122 122 N (observations) 917 917 917 Table A4 A4 Exponential event-history regression of democratic breakdown on aggregated support for democratic values, L2 predictors, and aggregated L1 controls (multilevel SEM approach) Model 3a Model 3b Model 3c b/se b/se b/se Intercept -5.547* -7.151 + -6.851 (0.563) (4.195) (4.332) Support for democratic values -0.911 + -1.009 + -0.945 (0.474) (0.592) (0.591) GDP 0.132 0.064 (0.428) (0.461) ELF 1.029 1.611 (2.141) (2.502) Age 0.696 (1.249) Education -0.644 (0.769) Subjective class 0.024 (0.949) AIC 42.444 46.179 51.203 BIC 52.086 65.463 84.951 N (failures) 5 5 5 N (subjects) 122 122 122 N (observations) 917 917 917 Table A5 A5 Exponential event-history regression of democratic breakdown on residualised support for democratic values (two-step approach) Model 4a Model 4b Model 4c b/se b/se b/se Intercept -5.460* -5.427* -5.433* (0.525) (0.517) (0.520) Residuals (Null model) -0.754 + (0.389) Residuals (model 1a) -0.658 + (0.357) Residuals (model 1b) -0.651 + (0.361) AIC 42.813 43.047 43.089 BIC 52.455 52.689 52.731 N (failures) 5 5 5 N (subjects) 122 122 122 N (observations) 917 917 917 In accordance with previous research, we use the terms 'conventional' or 'standard' multilevel analysis to describe hierarchical modeling techniques that are restricted to the analysis of level-one outcomes(Bennink, Croon, & Vermunt, 2013 Lüdtke et al., 2008; Preacher, Zyphur, & Zhang, 2010) . This points to the difference between formative and reflective models in measurement theory. Whereas formative latent variable models are already established in single-level measurement models(Diamantopoulos & Winklhofer, 2001) , it remains unresolved whether formative latent aggregation is equally possible. 4 Additional controls for measurement error can be integrated easily(Marsh et al., 2009) .For the sake of simplicity, our analysis of democratic persistence is limited to latent aggregation without controlling for measurement error. See Table A1(appendix) for a summary of all variables. 9 The latent aggregation was performed in Mplus, Version 7(Muthén & Muthén, 2012) . See Table A2 (appendix). Acknowledgments This paper benefited from feedback on several occasions including the SOCLIFE Graduate School at the University of Cologne, the annual conference of the European Survey Research Association (ESRA) in Lausanne and the joint conference of the Appendix The civic culture: Political attitudes and democracy in five nations G A Almond S Verba Princeton University Press Princeton, NJ Contextualizing cognitive consonance by a social mechanisms explanation: Moderators of selective exposure in media usage D Becker T Beckers S T Franzmann J Hagenah 10.1515/auk-2016-0108 Analyse & Kritik 38 1 Micro-macro multilevel analysis for discrete data: A latent variable approach and an application on personal network data M Bennink M A Croon J K Vermunt 10.1177/0049124113500479 Sociological Methods & Research 42 4 Stepwise latent class models for explaining group-level outcomes using discrete individual-level predictors M Bennink M A Croon J K Vermunt 10.1080/00273171.2015.1074879 Multivariate Behavioral Research 50 6 Political democracy and the timing of development K A Bollen American Sociological Review 44 4 Economic and noneconomic determinants of political democracy in the 1960s K A Bollen R W Jackman Research in political sociology R. G. Braungart & M. M. Braungart Greenwich, CT JAI Press Hierarchical linear models: Applications and data analysis methods A S Bryk S W Raudenbush Sage Newbury Park, CA Predicting group-level outcome variables from variables measured at the individual level: A latent variable multilevel model M A Croon M J P M Van Veldhoven 10.1037/1082-989X.12.1.45 Psychological Methods 12 1 Democratic challenges, democratic choices: The erosion of political support in advanced industrial democracies R J Dalton Oxford University Press Oxford, UK Index construction with formative indicators: An alternative to scale development A Diamantopoulos H M Winklhofer Journal of Marketing Research 38 2 A systems analysis of political life D Easton Wiley New York, NY A re-assessment of the concept of political support D Easton British Journal of Political Science 5 4 Changing mass attitudes and democratic deepening M D Fails H N Pierce 10.1177/1065912908327603 Political Research Quarterly 63 1 Ethnic and cultural diversity by country J D Fearon 10.1023/A:1024419522867 Journal of Economic Growth 8 2 The Oxford handbook of political behavior D Fuchs R. J. Dalton & H.-D. Klingemann Oxford University Press Oxford, UK Political culture paradigm Interaction between individuals and situations: Using HLM procedures to estimate reciprocal relationships M A Griffin 10.1016/S0149-2063(97 Journal of Management 23 6 Social mechanisms: An analytical approach to social theory Hedström, P., & Swedberg, R. Cambridge University Press Cambridge, UK Causal mechanisms in the social sciences. Annual Review of Sociology P Hedström P Ylikoski 10.1146/annurev.soc.012809.102632 36 Ethnic groups in conflict D L Horowitz University of California Press Berkeley, CA J J Hox Multilevel analysis: Techniques and applications New York, NY Routledge 2nd ed. Linking societal and psychological factors to homicide rates across nations F Lim M H Bond M K Bond 10.1177/0022022105278540 Journal of Cross-Cultural Psychology 36 5 Some social requisites of democracy: Economic development and political legitimacy S M Lipset 10.2307/1951731 The American Political Science Review 53 1 The multilevel latent covariate model: A new, more reliable approach to grouplevel effects in contextual studies O Lüdtke H W Marsh A Robitzsch U Trautwein T Asparouhov B O Muthén 10.1037/a0012869 Psychological Methods 13 3 Doubly-latent models of school contextual effects: Integrating multilevel and structural equation approaches to control measurement and sampling error H W Marsh O Lüdtke A Robitzsch U Trautwein T Asparouhov B O Muthén B Nagengast 10.1080/00273170903333665 Multivariate Behavioral Research 44 6 M G Marshall T R Gurr Polity IV Project Dataset users' manual M G Marshall T R Gurr K Jaggers Polity IV Project: Regime characteristics and transitions Vienna, Austria Center for Systemic Peace Civic culture and democracy: The question of causal relationships E N Muller M A Seligson 10.2307/2944800 American Political Science Review 88 3 Mplus User's Guide. Seventh Edition L K Muthén B O Muthén Muthén & Muthén Los Angeles, CA Critical citizens: Global support for democratic government P Norris Oxford Oxford University Press Social expenditure: Reference series Oecd 10.1787/socx-data-en Data file Modeling micro-macro relationships: Problems and solutions K.-D Opp 10.1080/0022250X.2010.532257 Journal of Mathematical Sociology 35 1-3 A general multilevel SEM framework for assessing multilevel mediation K J Preacher M J Zyphur Z Zhang 10.1037/a0020141 Psychological Methods 15 3 Democracy and the market: Political and economic reforms in Eastern Europe and Latin America A Przeworski Cambridge University Press Cambridge, UK Democracy and development: Political institutions and well-being in the world A Przeworski M E Alvarez J A Cheibub F Limongi Cambridge University Press Cambridge, UK Politics in plural societies: A theory of democratic instability A Rabushka K A Shepsle Merrill Columbus, OH Micro-macro links and microfoundations in sociology W Raub V Buskens M A L M Van Assen 10.1080/0022250X.2010.532263 The Journal of Mathematical Sociology 35 1-3 Democracy in divided societies: Electoral engineering for conflict management B Reilly Cambridge University Press Cambridge, UK The consolidation of democracy: Comparing Europe and Latin America C Q Schneider Routledge New York, NY A Latent Cluster-Mean Approach to the Contextual Effects Model With Missing Data Y Shin S W Raudenbush 10.3102/1076998609345252 Journal of Educational and Behavioral Statistics 35 1 Multilevel analysis: An introduction to basic and advanced multilevel modeling T A B Snijders R J Bosker Sage London, UK The quality of government standard dataset (Version Jan16) J Teorell S Dahlberg S Holmberg B Rothstein A Khomenko R Svensson 10.18157/QoGStdJan16 University of Gothenburg: The Quality of Government Institute Data file Explaining and understanding by answering "why" and "how" questions: A programmatic introduction to the special issue "social mechanisms U Tranow T Beckers D Becker 10.1515/auk-2016-0102 Analyse & Kritik 38 1 The changing face of methodological individualism L Udehn Annual Review of Sociology 28 World Values Survey 1981-2014 Longitudinal Aggregate (v. 20150418) World Values Survey Association. Data file (xsd:string)
?:linkScore	0.0 (xsd:float)
?:linkSource	GEFUREX pipeline (xsd:string)
?:linkingMethod	search (xsd:string)
?:mainEntity	<https://data.gesis.org/gesiskg/resource/study_program_90>
?:name	WVS (xsd:string)
is ?:referenceMetadata of	<https://data.gesis.org/gesiskg/resource/ssoar_reference_gesis-ssoar-57735_study_program_90>
?:toSource	gesis (xsd:string)
rdf:type	<https://data.gesis.org/gesiskg/schema/LinkMetadata>