The Correlates of Democratic Backslides

I merged ten databases grouping them by country and year standardizing the country names using the “countrycode” package, and manually re-coding deviating country names. The two core datasets are produced by the Varieties of Democracy (V-Dem) project: the main V-Dem v14 dataset and the Episodes of Regime Transformation (ERT) dataset. Third, the data on inequality comes from the World Inequality Database (WID). Compared to the others, this indicator has the lowest percentage of missing values compared to other known data and has the data for the earliest years. Nevertheless, even in this case most of the data was only present from the year 1985, which forbade me to compare post-WWII and post-Cold-War cases of democratic backsliding. The 26 percent of missing observations were filled with values predicted via bootstrapping accounting for the cross-sectional nature of data using the “Amelia” package in R based on several dozens of other socioeconomic indicators from the V-Dem Background Factors block. Fourth, I added the ParlGov dataset that approximates the ideological polarization between political parties in elections on the right-left scale. In that case, the database is limited to the most developed stable democracies and only has data for the election years. At the same time, its only direct competitor, Manifesto, has even more missing data. To access the data on societal affective polarization I used the dataset initially developed by Reiljan and later updated by Orhan that aggregates the existing survey data. Due to its origins, it is limited to 159 observations, two to three per country, which is devastating for the multilevel models. I combined the volatility indexes from five different regional datasets: one focused on the entire world with a particular focus on Latin America offered by Mainwaring and colleagues, Chiaramonte and Emanuele for Western Europe, Bertoa and Enyedi for geographic Europe, and Bogaards for Africa. In cases of overlaps between the datasets, the preference was given in the order they are mentioned. If the elections were regular, I carried volatility scores forward until the next elections. Table A3 on the data availability and the heatmap (Figure A1) are present in the online appendix. Even though the entrance of new parties into the playing field (known as extra-system volatility or type-B volatility) is theoretically more appealing than the overall volatility, these studies use different thresholds of what is considered to be the “system”. For instance, it varies from 2% for Mainwaring and Zoco and 1% for Chiaramonte and Emanuele.本研究将十个数据库按国家与年份进行合并，使用「countrycode」R包标准化国家名称，并手动修正存在偏差的国家名称。本研究的两个核心数据库均来自民主多样性（Varieties of Democracy, V-Dem）项目：主数据库为V-Dem v14数据集，以及政权转型事件（Episodes of Regime Transformation, ERT）数据集。第三，本研究的不平等指标数据来自世界不平等数据库（World Inequality Database, WID）。相较于其他指标，该指标的缺失值占比最低，且可追溯至最早的年份；即便如此，绝大多数数据仅可追溯至1985年，这导致本研究无法对二战后以及冷战后民主倒退的案例进行比较分析。本研究采用R语言的「Amelia」包，结合V-Dem背景因素模块中数十项其他社会经济指标，考虑到数据的截面特性，通过自助法（bootstrapping）填充了26%的缺失观测值。第四，本研究引入了ParlGov数据集，该数据集以左右维度为标尺，近似衡量选举中各政党间的意识形态极化程度。该数据库仅覆盖最发达的稳定民主国家，且仅包含选举年份的数据；其唯一的直接竞争数据集Manifesto的缺失值占比甚至更高。关于社会情感极化的相关数据，本研究使用了由Reiljan最初构建、后由Orhan更新的数据集，该数据集整合了现有调查数据。受原始数据来源限制，该数据集仅包含159条观测值，每个国家约2至3条，这对多层模型（multilevel models）的构建极为不利。本研究整合了五个不同区域数据集的波动率指数：一是Mainwaring及其团队构建的、聚焦全球尤其是拉丁美洲的数据集；二是Chiaramonte与Emanuele针对西欧的数据集；三是Bertoa与Enyedi针对欧洲地理区域的数据集；四是Bogaards针对非洲的数据集。当数据集存在重叠时，将按照上述提及的顺序优先选用对应数据。若选举为定期举行，则将波动率得分沿用至下一次选举。本研究的数据可用性表（表A3）与热度图（图A1）详见在线附录。尽管政党新进入选举场域（即体系外波动或B类波动）在理论上比整体波动率更具研究价值，但现有研究对「体系」的界定阈值各不相同：例如Mainwaring与Zoco设定为2%，而Chiaramonte与Emanuele则设定为1%。