ReSOG’s contents

ReSOG comprises twelve original variables, all included in V-Dem’s ‘Regimes survey’. The variables are listed alongside their country coverage, time series, and aggregation method (from expert to country level) in Table 1. Five variables pertain to regime support groups, six to regime opposition groups, and one scores the most powerful social group in a country. Four support group variables (Regime support groups, Most important support group, Support group size, Support group main location) were part of the Historical V-Dem data collection (Knutsen et al. Reference Knutsen, Teorell, Wig, Cornell, Gerring, Gjerløw, Skaaning, Ziblatt, Marquardt, Pemstein and Seim2019). All twelve ReSOG variables are included in the latest versions of the contemporary V-Dem coding, which started in 1900. Hence, the maximum time series in V-Dem v.12 (Coppedge et al. Reference Coppedge, Gerring, Knutsen, Lindberg, Teorell, Alizada, Altman, Bernhard, Cornell, Fish, Gastaldi, Gjerløw, Glynn, Grahn, Hicken, Hindle, Ilchenko, Kinzelbach, Krusel, Marquardt, McMann, Mechkova, Medzihorsky, Paxton, Pemstein, Pernes, Ryden, von Romer, Seim, Sigman, Skaaning, Staton, Sundstrom, Tzelgov, Wang, Wig, Wilson and Ziblatt2022a) extends from 1789–2020 for four variables, and eight variables are coded from 1900–2020.

Table 1.

Variables included in ReSOG

Note: See V-Dem Codebook, v.12, pp.138–147 for details on question clarifications, answer categories, measurement level, and variable-specific rules for dealing with ‘ties’ (that is, two modal categories).

Let us dwell on the fourteen social group answer categories listed in Table 2, which recur for eight ReSOG variables on the identity of (most important) support groups, opposition groups, or most powerful groups. For illustration, Figure 1 maps the most important support group variable, globally, in 2000. The categorization scheme is intentionally multi-dimensional (as described to expert coders) with potentially overlapping categories, aiming to capture several relevant types of social groups emphasized by different literatures. Dimensions pertain, for example, to the urban-rural divide and class divisions, but we also include categories to capture historically salient groups occupying key roles in the state (military and civil servants) and groups primarily described by their ethnic or religious affinity. Hence, we prioritized comprehensiveness and capturing several relevant groups (in different contexts) rather than having a categorization scheme that follows one organizing principle with mutually exclusive categories.

Table 2.

ReSOG’s fourteen social group categories

Figure 1.

Mapping the most important support group variable (v2regimpgroup) in 2000.

Moreover, we do not require that these groups act as unitary actors for being registered as regime support or opposition groups. For example, in our description to expert coders for regime opposition groups, we note that experts shall register those groups that ‘include noteworthy opposition actors – that is, individuals (mobilized or not) who both want to and who could, under favorable circumstances, be able to remove the existing political regime.’ Hence, a particular group (say ‘the military’) might register both as a regime support group (for example, army officers supporting the current regime and its junta) and an opposition group (for example, naval officers planning a coup to replace the current junta with another one).

Collecting and processing the data

Around 1,000 country experts have coded ReSOG variables, whereof more than 100 are historical experts, typically political historians or political scientists who have researched eighteenth-century politics for the country in question. Most ‘contemporary country experts’ have coded their country from 1900–2020, whereas historical experts coded their polities from the start of the time series to 1920 (to ensure twenty years of overlap with contemporary experts). For additional details, see Knutsen et al. (Reference Knutsen, Teorell, Wig, Cornell, Gerring, Gjerløw, Skaaning, Ziblatt, Marquardt, Pemstein and Seim2019) on historical coders and Coppedge et al. (Reference Coppedge, Gerring, Knutsen, Lindberg, Teorell, Marquardt, Medzihorsky, Pemstein, Fox, Gastaldi, Pernes, Rydén, von Römer, Tzelgov, Wang and Wilson2022) for post-1900 coding. The ambition is to have at least five experts per country question for the 1900–2020 period.Footnote ⁵ Since there are fewer potential historical experts in most countries, the protocol for historical coding sets additional requirements and quality controls for recruitment and higher remuneration for additional coding time, to compensate for the fewer experts.

We take several measures to clarify the relevant concepts and (regime) units to be coded. This is especially important since we want to measure complicated concepts and because partially overlapping concepts referred to by different, but resembling, terms exist (such as leader’s ruling coalition or winning coalition). Hence, we cannot guarantee that all experts have the same reference in mind when coding; some might misunderstand and code; for example, the leader’s ruling coalition rather than the regime’s support coalition, as we define it. This could result in measurement error and higher uncertainty for some countries and regimes. For instance, if differences between the leader’s ruling coalition and the regime’s support coalition are smaller in personalist dictatorships than in democracies, and some experts confuse the concepts, the induced errors are larger for democracies. While we cannot remove such errors entirely in a project where about 1,000 experts code, we do present all experts with initial clarifications of the underlying units (regimes) and core concepts (regime support and opposition groups) once they start the online survey. Moreover, all questions come with explicit definitions of key terms, detailed clarifications for questions and answer categories, and even illustrative examples. We refer readers to the V-Dem codebook (Coppedge et al. Reference Coppedge, Gerring, Knutsen, Knutsen, Lindberg, Teorell, Altman, Bernhard, Cornell, Fish, Gastaldi, Gjerløw, Glynn, Grahn, Hicken, Kinzelbach, Marquardt, McMann, Mechkova, Paxton, Pemstein, von Römer, Seim, Sigman, Skaaning, Staton, Tzelgov, Uberti, Wang, Wig and Ziblatt2022, 138–147; see also Appendix I), which contains the same definitions, clarifications, and examples that experts view when coding.

Perhaps most importantly, we display, for all experts, the pre-coded data on the identity and start and end dates of the regime as well as the process through which the regime ended (see Djuve et al. Reference Djuve, Knutsen and Wig2020 for a presentation of these pre-coded data on more than 2,000 regimes). We do so to ensure that all experts know what we regard as the regime unit that should be considered when they code characteristics of regime support and opposition groups. This clarification is especially important for ensuring consistency during times of civil war or other types of divided territorial control. It also serves to remind experts of, and concretize, the more abstract point that we – in democracies – are not considering the supporters or opponents of the specific government, but instead which groups support/oppose and are important for maintaining/replacing the wider democratic system. This point on how to code democracies is also made explicit to coders in separate question clarifications.

Yet, since experts will inevitably disagree on the relevant support and opposition groups in many cases, even if they accept our definition of what the regime unit is, we need rules for aggregating across coders. The aggregation methods for translating expert scores into country scores, listed in Table 1, are question-specific. For example, when asking experts about the most important opposition group, we opt for the modal category – that is, the category chosen by most country experts for a country year – with ties being decided by which category has the highest score on the regime opposition group question. For this latter question, which is a multiple selection question with fourteen social group categories, we aggregate by taking the mean across experts. To exemplify, if there are five experts, and three consider the urban working class to be an opposition group and two experts do not, the country year aggregate will be (3*1 + 0*2)/5=0.6. In the next section, we discuss how to interpret this aggregate score and discuss transformations of the original measures.

For the regime support and opposition groups size variables, which are the only two variables originally at an ordinal measurement level, with experts scoring them on five-category scales, we use V-Dem’s measurement model for aggregation (see Pemstein et al. Reference Pemstein, Marquardt, Tzelgov, Wang, Medzihorsky, Krusell, Miri and von Römer2022; Coppedge et al. Reference Coppedge, Gerring, Glynn, Knutsen, Lindberg, Pemstein, Seim, Skaaning and Teorell2020). This Bayesian IRT model draws on different pieces of information, including experts coding hypothetical cases in anchoring vignettes, to anchor expert scores and make them more comparable. The model ultimately produces country-time estimates on a latent interval scale along with uncertainty estimates.

ReSOG’s variables contain rich information about political systems across countries and times that should be of interest to social scientists studying various topics. Notably, this goes also for derivative measures that may easily be constructed from the twelve original variables. To exemplify, we can easily construct derivative measures of so-called ‘cross-class coalitions’ (see, for example, Esping-Andersen Reference Esping-Andersen1990; Luebbert Reference Luebbert1991), to capture coalitions (be it as regime opponents or in the support coalition) that include actors from two or more economic classes. Still, we should note that many group categories are rather broad and may miss more subtle changes in the coalitions supporting or opposing regimes. Take ‘ethnic or racial group(s)’; a regime could initially be supported by several, large ethnic groups and later lose support from most of them. While our support coalition size measure could capture an associated reduction in numbers, the specific support group dummy would regardless be coded ‘1’.Footnote ⁶ Hence, ReSOG’s measures alone cannot give a complete description of the coalitions supporting or opposing regimes, and using information also from complementary datasets could enrich descriptions. For example, EPR data could capture additional nuances in coalition developments by tracking changes in supporting ethnic groups.

Validation and reliability assessments

We ran convergent validity tests using other data sources. The first is the ongoing Paths to Power data collection by Nyrup et al. (Reference Nyrup, Knutsen and Langsæther2023) on the occupational, educational, and social background of cabinet ministers – a small but important subset of support coalition members – globally after 1966. In Appendix III, we use aggregated versions of their individual-level measures for comparisons against relevant ReSOG regime support group measures, typically finding moderately strong, positive correlations.

Second, we already discussed Schultz and Kelsall’s (Reference Schulz and Kelsall2021) rich dataset, and one variable well suited for comparing against ReSOG’s regime support coalition size measure is Leader Block as a percentage of the population (LB per cent). Yet, the latter measures the size of the coalition behind the leader rather than the regime. Moreover, it differs from our measure by asking coders to only assess who supports the regime rather than requiring both support and that this support is effective in enhancing the leader’s/regime’s grip on power. Thus, ReSOG’s measure is more demanding, and we, therefore, expect typically lower scores on the regime support coalition size measure than on LB per cent. Given these conceptual differences, we also expect far from a perfect correlation, but a modestly strong positive one.

Indeed, this is what we find, and, reassuringly, the correlation is quite persistent across time for around 30–40 countries with data on both measures. Figure 2 shows scatterplots with best-fit lines for 1960 (r=.44), 1980 (r=.50), 2000 (r=.52), and 2018 (r=.56), the last year of Schultz and Kelsall’s data. Overall, the fit between the two measures is about as we expected – fairly strong but with several outliers. Notably, several countries with intermediately sized support coalitions have quite high LB scores. These are (presumably) autocracies with popular leaders (hence the high LB per cent scores), albeit with a smaller set of powerful backers that are key for keeping the regime afloat. One example is the single-party, majority Hutu-dominated Habyarimana regime in 1980-Rwanda, which had an LB score of 62 per cent. Yet, the regime received a middling 0.40 support group size score (the 1980-median score, globally, was 0.77), indicating that the key supporters effectively keeping the regime in power were fewer than the popular majority. (The support groups coding indicates that party elites and military officers were among the support groups.) Another example is Museveni’s electoral authoritarian regime in 2000-Uganda; Museveni was re-elected with 69 per cent of votes in 2001, close to his LB score of 63 per cent. Yet, the support group size score of 0.94 was below the median global score (1.20) that year, suggesting a more concentrated set of core supporters maintaining the regime in power.

Figure 2.

Convergent validation: Scatterplots with best-fit lines for four selected years for our regime support coalition size measure (x-axis) and Leader Block as a percentage of the population from Schultz and Kelsall (y-axis).

Third, assuming that there is typically more consistency in the presence of particular regime support groups during a regime’s tenure than before and after regime changes, we use data on regime units from Geddes et al. (Reference Geddes, Wright and Frantz2014) and groups of connected leaders (with similar coalitions) from CHISOLS (Mattes et al. Reference Mattes, Leeds and Matsumura2016) for validation. As expected, we find substantially more between-regime than within-regime variation in regime support groups. We refer to Appendix IV for results. Here we also separately display within- and between-regime variance for autocracies, democracies, and very stable democracies. In Appendix X, we present country examples showing both between- and within-regime developments in regime support groups size and support group identities.

In the following, we consider variations in and estimated reliability of coder-level scores. Yet, before doing so, we highlight that the uncertainty pertaining to support and opposition group characteristics, as indicated by varying degrees of expert agreement, is an important part of the knowledge that we produce and report. Take, for example, ReSOG’s regime support group measure, and consider the ‘military officers’ category. Experts are in full agreement, for example, that the military was part of the regime support coalition in Myanmar in 2002 (score 1) or that it was not part of it in Norway in 2016 (score 0). Experts were, however, not in full agreement on whether Mobutu’s regime in Zaïre in 1980 included the military in the support coalition, though the 0.75 score means that three of four experts held this belief. For yet other countries, such as France-1946 or Saudi Arabia-1997, experts were equally split (0.5 scores) on whether or not the military was in the regime support coalition.Footnote ⁷ Such middling scores may not necessarily reflect poor validity – they may reflect an intermediate degree of political influence for the group (here military), with experts having different thresholds for registering groups as sufficiently important for maintaining regime stability.Footnote ⁸ Yet, at the very least, expert disagreement tells us something about the certainty with which we can make strong claims on the presence or absence of these often hard-to-observe features of regime support coalitions.Footnote ⁹

To further assess measurement validity and reliability, we explore, more systematically, the correlates of uncertainty and expert-coder disagreement, focusing on the regime support coalition size variable. First, we follow McMann et al. (Reference McMann, Pemstein, Seim, Teorell and Lindberg2022, 438) and ‘assess systematic determinants of respondent disagreement in a regression framework […] where the dependent variable is the standard deviation of measurement model-adjusted ratings among respondents for each country and year’. These regressions, presented in Appendix Table A1, include up to 25,975 country-year observations, and we assess different covariates. First, we include the number of coders for each observation as a covariate; expectedly, uncertainty in country-year aggregate scores drops as this number increases. Second, we include support coalition size (linear and squared terms), as size might be harder to code precisely for smaller, intermediate, or larger coalitions. Next, we include a linear time trend and a dummy for whether the observation is coded by Historical V-Dem, anticipating that coalitions of the distant past are harder to code precisely. We also include geographic region-fixed effects and ln GDP p.c. (from Fariss et al. Reference Fariss, Anders, Markowitz and Barnum2022), expecting that coalition size is harder to code in poorer countries. Further, we include a democracy dummy from Skaaning et al. (Reference Skaaning, Gerring and Bartusevičius2015) and ln population, impartial administration, and freedom of expression from V-Dem. We anticipate that smaller countries with partial administrations and poor protection of free speech are harder to code accurately.

Some of these factors correlate with the uncertainty measure. Notably, smaller countries have more uncertain codings (Appendix Table A1), as one might expect if smaller countries are less well-documented in various sources. However, further scrutiny suggests that much of this association comes from smaller countries having fewer V-Dem expert coders; in multivariate regressions, we find no robust direct effect of population size. Indeed, there aren’t too many robust correlates of uncertainty in support coalition size scores, especially once controlling for alternative factors. Uncertainty does not depend, systematically, on income, impartial administration, free expression, or democracy.Footnote ¹⁰ Surprisingly, there is no robust relationship with the year being coded either, although the bivariate correlation is sizable and the negative time trend coefficient indicates that recent years have more certain codings. The lack of strong correlations is reassuring since their presence could have suggested systematic biases with lower-quality coding of particular groups of observations. Yet, in specifications including a squared coalition size term, we find clear indications that intermediately sized coalitions are associated with less uncertainty than larger and smaller ones. Also, there are some indications of regional differences, with Latin American coalitions having higher uncertainty than Eastern European ones. (This pattern is not replicated for Opposition group size). Finally, the strongest pattern is that having more expert coders per observation systematically reduces uncertainty. As displayed in Figure A.3, the median number of experts per observation has increased to five in V-Dem’s v.12, thereby enhancing reliability relative to previous releases.

Varying expert numbers also explain differences in measurement errors for other ReSOG variables (Appendices VIII-IX). Hence, dataset users should note that there are generally fewer expert coders for the regime opposition group variables than for the support group ones (coded also for V-Dem v.9-v.10). Overall, reliability and validity may therefore be higher for the latter ReSOG variables. We also highlight that users wanting to, for example, describe country-specific trends or run regressions should inspect accompanying information on expert coder numbers and consider excluding post-1900 observations (at least in robustness tests) with, say, <4 coders. Appendix VIII reproduces descriptive figures from the next section after making this restriction, but the global trends that we consider remain fairly similar.

Next, we shift focus from the country-year level to the country-year-coder level, treating every expert score as an observation. This allows us to assess not only the relevance of country-level factors but also expert-level characteristics for reliability. The tests in Table A2 use different measures of expert-level reliability in the regime support coalition size coding. These include the i) absolute distance between the expert’s score and the mean score for that country year, ii) so-called beta scores from V-Dem’s measurement model, with high scores implying high expert reliability for the particular variable, and iii) experts’ own (0–100) confidence rating in their scores (which can vary by year). In addition to including most country-level measures from above, we add information from V-Dem’s post-questionnaire survey on experts’ sympathy with two core principles of democracy (electoral and liberal), reported primary motivation for coding, satisfaction with the V-Dem coding process, and time spent on coding.

Few country-level factors are systematically related to the expert-level uncertainty measures. The exception is higher reliability in democracies than autocracies for all three outcome measures. Among the coder-level covariates, (self-reported) time spent on the coding task is not related to reliability. Sympathy with principles of democracy is mostly insignificant as well. By contrast, experts who were more satisfied with the coding experience provided scores that deviated less from the mean and obtained higher (reliability) beta scores. Finally, the main motivation for coding is correlated with reliability. Experts who report that ‘being a part of the V-Dem network provides benefits for my reputation’ or that V-Dem data ‘is a valuable tool for scholars and policy makers’ are generally more reliable, according to different metrics, than those motivated by money (reference category).

Finally, one reliability issue resulting from expert disagreement requires attention: in some instances, scores may change for countries from one year to the next even if there are no political changes, but due to the sets of experts being dissimilar for the two years. Such changes are particularly likely in 1900 (due to the inclusion of several contemporary experts) and to a lesser extent in 1920 (the last year of coding for the fewer historical experts). Another issue pertains to country-level heterogeneity. It might, for example, be that experts coding largely stable democracies apply different thresholds than those coding countries with more varied or autocratic regime histories when coding, say, which groups could ‘under hypothetical ‘favorable conditions’, be capable of removing the regime’ and thus qualify as regime opposition groups. Dataset users should be aware of these features when presenting and interpreting descriptive statistics, and we strongly advise accounting for breaks in the time series. We also strongly advise accounting for temporal and country-level heterogeneity (for example, via year- and country-fixed effects) when using ReSOG data in regression analysis.

The size and heterogeneity of support and opposition groups

We begin by considering one theoretically interesting measure that can be constructed from the original data pertaining to the support coalition’s heterogeneity. This heterogeneity measure is straightforward: It relies on dichotomized versions of the fourteen support group variables – each scored 1 if half or more of country experts agree the group in question is a support group (that is, when the original score for the support group category in question is ≥0.5). For the final heterogeneity measure, we then sum up the number of support groups across these fourteen derived dummies for each country-year observation.Footnote ¹¹

Figure A3 (top plots) shows histograms for the number of support groups for 6,454 democratic- and 18,637 autocratic observations. The democracy-autocracy categorization is based on a dummy variable scored 1 (democratic) if the regime scores ≥4 on Skaaning et al.’s (Reference Skaaning, Gerring and Bartusevičius2015) Lexical Index of Electoral Democracy (v.6.3), implying that democracies are regimes with contested multi-party elections. The histograms reveal that democratic regime support coalitions include more social groups than autocratic ones when averaging across the entire historical period. The modal number of support groups is four for democracies and one for autocracies, whereas their respective medians are five and three.

Yet, the histograms also reveal considerable variation in support coalition heterogeneity within the regime types. Of the autocracies, 16 per cent have more than five support groups – examples being Mexico 1930–1988 under the PRI and Austria 1848–1918 under the Habsburg Monarchy. Some democracies are listed with few support groups. For example, late 1980s Thailand is only registered with the state bureaucracy and military as support groups, and Botswana 1966–1979Footnote ¹² with only agrarian elites and business elites (although rural working classes and party elites came very close to reaching the 0.5-threshold, with 0.43 scores for both groups).Footnote ¹³

Let us now turn to the measures on regime support groups size and opposition groups size, originally coded on ordinal scales ranging from 0, ‘Extremely small (About 1 per cent of the population or less)’ to 4, ‘Large (More than 30 per cent’). The country-expert scores are subsequently aggregated and transformed into latent interval scales by V-Dem’s measurement model. The resulting variables extend empirically from -3.9 to +2.9 (support groups size) and -3.8 to +2.1 (opposition groups size). The Kernell density plots in Figure 3 display distributions of scores for all coded observations, broken down by regime type.

Figure 3.

Kernell density functions: Distribution of regime support groups size (left) and regime opposition groups size (right), by regime type. All years and countries included.

Figure 3 (left plot) shows that, in general, regime support groups size is higher in democracies than autocracies, with means of, respectively, 1.4 and -0.6, and medians of 1.6 and -0.6. Further, regime opposition groups size is generally lower in democracies than in autocracies. For regime opposition groups size (right plot), the variation is relatively large and not too different between the two regime categories. For support groups size, however, autocracies display a much larger variation, with a variance of 1.8, compared to 0.8 for democracies (the latter variation stems, inter alia, from including several low-suffrage competitive regimes under the minimalist democracy definition).Footnote ¹⁴ Whereas 6 per cent of autocracies have higher scores than the democratic median of 1.6 – and a few autocracies such as Putin’s Russia from 2004–2008 and Malaysia from 1958–2017 have scores exceeding 2 – numerous autocracies have very low scores.

There are plausible reasons why autocracies vary so much in regime support coalition size. As highlighted in the autocratic politics literature, autocracies are very heterogeneous in terms of how power is organized, who wields power, and in linkages between the regime and different population groups (for example, Svolik Reference Svolik2012; Geddes et al. Reference Geddes, Wright and Frantz2018). These insights have spurred several efforts to categorize autocracies into types (for example, Wahman et al. Reference Wahmann, Hadenius and Teorell2013; Geddes et al. Reference Geddes, Wright and Frantz2014) and the resulting regime categorizations have even explicitly been used as proxies for differences in the size of coalitions supporting the leader (Bueno de Mesquita et al Reference Bueno de Mesquita, Smith, Siverson and Morrow2003; Mattes et al. Reference Mattes, Leeds and Matsumura2016). For example, it is widely assumed that dominant-party autocracies have much broader coalitions than military regimes. With our new support coalition size measure, we can assess this assumption empirically.

Indeed, when considering regime support groups size scores across Geddes et al.’s categories – as reconstructed by Anckar and Fredriksson (Reference Anckar and Fredriksson C2019), who code autocracy type back to 1800 – we find evidence that dominant party regimes have much larger regime support coalitions than the other autocracy types. The mean score for dominant party autocracies is 0.9 whereas the mean score is 1.4 for democracies and -0.2 for all other autocracies. Among the other autocracy types, personalist regimes have generally higher scores (mean of 0.0) than military regimes (-0.3) and autocratic monarchies (-0.3). Hence, the broad patterns in our data follow commonly invoked assumptions about differences in coalition size across different autocracy types.Footnote ¹⁵ Also when separating closed and electoral autocracies, we find that electoral autocracies have larger (and more heterogeneous) regime support coalitions.Footnote ¹⁶

However, Figure 4 shows that assumptions tying different autocracy types to differences in coalition size are simplifying. For all four autocracy categories, there is considerable variation. For instance, more than 10 per cent of military regime observations have higher support groups size scores than the median dominant party regime – the military regime in Egypt under Nasser in the 1950s and 60s, for example, had a support groups size score of about 1.3. This score might be indicative of decolonization being associated with the emergence of relatively inclusionary autocracies that sprung up globally in opposition to imperialism.Footnote ¹⁷ By contrast, various nineteenth-century military regimes in Bolivia operated with very narrow coalitions, scoring well below -2 for example in the 1820s–30s. For dominant party regimes, one small-coalition regime is the Afghan one during the Soviet invasion in the 1980s (scores around -1.1), whereas a quintessential large-coalition regime is China in the 2000s (around 1.7). Hence, our data suggest that researchers should take caution before using existing autocratic regime-type categorizations to proxy for coalition size.Footnote ¹⁸

Figure 4.

Distribution of regime support groups size (top) and regime opposition groups size (bottom), by autocratic regime type. Kernell density functions. Only countries coded as one of the autocracy types (back to 1800) by Anckar and Fredriksson are included.

As for opposition groups size, Figure 4 (bottom graph) shows a similar picture of general differences between the regime type categories. The largest mean opposition groups size score is for military regimes – incidentally, also the shortest-lived regime type (Geddes et al. Reference Geddes, Wright and Frantz2014) – and the smallest mean is registered for autocratic monarchies – incidentally, the longest-lived autocracies. Yet, as for support coalition size, there is vast within-category variation.

Further assessments show that at any point in modern history, there have been notable cross-country differences in support coalition features. For support coalition size, the between-country variance (1.2) is only slightly lower than the over-time variance within countries (1.3). For example, in 1790, France had a support groups size score of about 0, and its regime support groups (according to our dichotomized variables) included Agrarian elites, Business elites, Urban middle classes, and Rural middle classes. In the Ottoman Empire in the same year, the support groups size score was -2.1, and only the State bureaucracy reached the threshold for registering as a support group. In 2020, democratic Belgium was among the countries with the largest number of support groups (10 out of 14), whereas autocratic Eritrea (Military) and Turkmenistan (Local elites) registered only one each.

The ReSOG data also reveals notable historical trends. There is a substantial increase in average regime support coalition size over time. For example, the mean support groups size score was -1.1 among the sixty countries with data in 1789, whereas the corresponding number, for 176 countries with data, was 1.1 in 2020. Figure 5 (top graph) shows that the increase was gradual but consistent from 1789–1900, and the drop in the global average in 1900 is related to the expanded sample including about fifty African and Asian colonies with generally small regime support coalitions. The increasing trend resumed after WWI and continued at a much faster pace after the Second World War to the 1980s. After 1990, the mean support group size has basically stagnated.Footnote ¹⁹

Figure 5.

Global averages of regime support- and opposition groups size. Top: Means for all countries with data. Bottom: Means by regime type, using a dummy based on LIED (contested elections) to distinguish democracies from autocracies. Vertical line for 1900 marks year with the expansion of V-Dem sample to include approximately fifty African and Asian colonies.

Further investigation of the increase from 1900–1990 indicates there were two main contributing factors to the trend.Footnote ²⁰ First, the large increase in the share of democracies globally, from around 11 per cent in 1900 to around 46 per cent in 1990, contributed substantially. Democracies have had considerably larger coalitions than autocracies across the entire time period, especially prior to 1960. In 1900, for example, the mean score for democracies was around +1 and the mean score for autocracies was around -1. In 2020, the difference had shrunk from 2.1 to 1.2 points but was still considerable.

Second, Figure 5 (bottom) shows that the average support coalition’s size in autocratic countries trended considerably upwards, especially from around 1920 to the 1980s. The mean score in autocracies was -1.1 in 1900 and 0.4 in 1980. This trend has corresponded with several notable developments in autocracies, including the increased spread of multi-party elections (Miller Reference Miller2015) as well as parliaments and regime parties (for example, Frantz Reference Frantz2018). Interestingly, the mean number of support groups in autocracies has not changed considerably during the same period (Appendix Figure A.4).

The historical rise of some regime support groups and the decline of others

We now turn to charting the social identity of regime support groups over time. We find clear indications that the types of social groups that regimes have drawn support from to stay in power have changed during modern history, with the replacement of certain (less numerous) elite groups in several countries by more populous social groups such as the urban working classes. We will discuss how much of this trend is related to regimes becoming more democratic. Yet, also some consistently autocratic countries experienced such changes. One notable example is Russia/the Soviet Union, which has remained autocratic throughout the entire period. Still, the Tsarist regime, which according to our data had a support groups size score of 0.2 in 1900 and relied on support from the aristocracy, agrarian elites, the state bureaucracy, and the military, was replaced, eventually, by the Bolshevik regime. Throughout the 1960s, for example, the latter regime had a support coalition size of 1.8 and counted party elites, the state bureaucracy, the military, urban working classes, and rural working classes among its support groups. Not only had the number of regime supporters expanded, but the coalition’s composition had changed so that old elite groups mainly located in the countryside had been exchanged with non-elite groups, including the working classes residing in the cities.

The demise of these specific (rural) elite groups as core regime supporters is not restricted to Russia. Figure 6 shows that it is a global phenomenon and that the gradual demise of the aristocracy and agrarian elites has coincided with the historical rise to power of major urban groups. The figure shows time trends in the global means of the (dichotomized) regime support group categories for the aristocracy, agrarian elites, urban middle classes, and urban working classes. About 75 per cent of the regimes at the time of the French Revolution included the Aristocracy in their support coalitions, but this share rapidly declined to around 50 per cent in 1848, the ‘Year of Revolution’. With the counter-revolution removing several liberal regimes already from 1849, the share temporarily bounced back in the 1850s, before once again descending from the 1860s. It gradually declined to below 10 per cent in the 1980s, where it has stayed since.

Figure 6.

Share of regimes, globally, with the group included in the support coalition, based on dichotomized support group measures. The vertical line marks the year (1900) with a large expansion of the V-Dem sample to include approximately fifty African and Asian colonies.

Agrarian elites followed a slightly different trajectory, with an increasing trend from 1789 (included in around 50 per cent of support coalitions) until the 1880s (around 60 per cent). This increase is mainly driven by developments in Latin American countries, where agrarian elites became even more prominent powerbrokers during the nineteenth century. From the 1880s onwards, however, agrarian elites have become increasingly rare as support coalition members (the drop in 1900 is mainly due to the changing sample, with around fifty Asian and African colonies added). Over the last forty years, this rural elite group has entered in fewer than 10 per cent of regime support coalitions, globally.

Two urban non-elite groups, namely the urban middle and urban working classes, have trended in the opposite direction. Presumably, these groups’ rise to power has been driven by the urbanization and industrialization of economies from the late eighteenth century and the democratization of many political systems (often achieved exactly through these two groups revolting; see, for example, Dahlum et al. Reference Dahlum, Knutsen and Wig2019). In 1789, fewer than 10 per cent of support coalitions included the urban working classes, and the same applied to the urban middle classes. By 2020, around 25 per cent of the world’s regime support coalitions included urban workers and around 40 per cent of the urban middle classes.

Still, there are some interesting differences between these two groups’ trajectories: The urban middle classes started being incorporated as regime support groups earlier than the urban working classes; the percentage share of countries where the urban middle classes were included increased from below 10 per cent in 1789 to about 35 per cent in 1899. (The drop in 1900 is once more related to the expanded sample, which was dominated by Western European and American countries pre-1900). By contrast, the urban working class share only increased by around 5 per cent from 1789–1899, with the single notable short-term increase coming in 1848.

From 1900 to the mid-1980s, however, the increase was relatively larger for the urban working classes, leading to the two groups’ global means being relatively close towards the end of the Cold War. Thereafter, the global mean for the urban working classes stagnated, whereas the urban middle classes were included in the support coalitions in numerous countries in the 1990s. Around 2000, the urban middle classes were present in about twice as many support coalitions as the urban working classes, and while their global means have converged a little recently, most of the difference has been retained.

Behind these global trends, there is considerable cross-regional heterogeneity. Figures A8-A13, depicting these regional trends, are presented and discussed in Appendix VI. Notably, the trends of aristocratic and agrarian elite decline as well as the rise to power of the urban middle and working classes started earlier in some regions, especially Western Europe and North America, than in others. In the MENA region the aristocracy scores in the twenty-first century are almost similar to those in the nineteenth century.