Modeling the Belarusian Polity

Using predictive modeling methodologies to model the Belarusian polity with open source survey data.

Introduction

There are currently unknown quantities of Wagner personnel stationed in Belarus, engaging with Belarusian forces in military exercises and causing significant concern for neighboring Latvia, Poland, and Lithuania. Belarusian President Lukashenko, sometimes regarded as 'the last dictator in Europe,' is openly allied with Putin and supports Putin's political objectives by hosting Russian nuclear launchers. Like Ukraine, Belarus is a former Soviet republic that Putin fears may move toward European integration and western alignment, isolating Russia and potentially causing democratic spillover effects that could undermine his autocratic rule. Despite Lukashenko's regime fostering near-total military and political subordination of Belarus to Russia, there remain Belarusian democratic opposition movements that seek to affirm Belarusian sovereignty from Putin's influence. These democratic hopefuls face significant challenges, including state-funded influence operations aimed at maintaining and expanding Belarusian allegiance to Russia (Mankoff, 2024).

Effective political messaging targeting is crucial in supporting Belarusian democratization efforts. Identifying the right audience for pro-democratization messages involves understanding the complex dynamics of the Belarusian polity. Traditionally, this task has relied on qualitative analysis of open-source information. However, modern approaches require robust political science research that thoroughly analyzes and models the Belarusian population to craft messages that resonate with key demographic groups.

This research seeks to answer the question, “What are the attributes of a promising target audience in Belarus for pro-democratization political messaging?” It aims to provide insights into the makeup of the Belarusian polity to inform strategies for delivering effective political messages. Utilizing a recent Belarusian survey dataset conducted by Chatham House, this research employs hierarchical clustering and a classification tree to identify relevant insights into the Belarusian population. The study identifies a particular cluster of Belarusians as promising targets for pro-democratization messages and highlights key attributes of Belarusians who are open to or resistant to European integration.

Data 

This research utilizes the Eighteenth Survey Wave of the Chatham House Belarus polling, titled “Belarus: The Changing Social Contract,” conducted from February 3rd to February 22nd, 2024. The survey included 837 respondents and was conducted using a computer-assisted web interview method (Eighteenth Survey Wave, n.d.). The study notes that the survey may underrepresent support for Lukashenko, as his supporters tend to be less socially and politically active and less likely to participate in web-assisted surveys. Additionally, the "fear factor" of living under authoritarianism may cause the survey to underrepresent Belarusians critical of the current government.

Translation Efforts

Chatham House posted their survey data in XLSX format in Russian, necessitating programmatic translation. I used pandas and Google Translator from the deep_translator package to translate the dataset. Initializing the translator with the source language 'ru' (Russian) and target language 'en' (English), I translated the column names of the DataFrame. This process iterates through each column name, translating it using the Google Translator API, and stores the translated names in a list. A dictionary mapping original column names to their translated counterparts was created, and the DataFrame's columns were renamed accordingly. The translated DataFrame was saved to a CSV, and a function, translate_values(), was used to translate the unique values in each column while skipping specific values like "#NULL," "Checked," "Unchecked," and "nan." The translated values were organized into a dictionary, the dictionary was then used to translate the entire dataset. This dictionary mapping approach utilizing the unique values was necessary due to the dataset's size, which precluded simple translation of each field due to time constraints and incessant API error obstacles.

Methods

Clustering

I first applied hierarchical clustering to the transformed dataset. This approach was preferable to K-means clustering due to the lack of intuition around an optimal number of clusters for modeling the Belarusian polity. Mean data imputation was performed to fill NA values in the dataset. The hierarchical clustering algorithm used Euclidean distance between rows, treating each row as its own cluster and iteratively comparing pairs to identify the most similar clusters based on Euclidean distance. These most similar clusters were fused in a dendrogram, with observation similarity represented by lower fusion points along the Y-axis. After generating the dendrogram, I cut it at four clusters.

Tree-Based Modeling

A tree-based classification model was used to predict the variable anti_eu_integration, which indicates opposition to Belarus becoming an EU member state as a proxy for preferring non-Western political alignment. This analysis assumes that opposition to European integration correlates with preference for or susceptibility to Russian political alignment. While this is an imperfect assumption, it is useful given the dataset's limitations. I split the dataset into a test and training set, built a decision tree model using the tree() function, and predicted the anti_eu_membership variable based on a set of predictor variables (total_monthly_income, education_level, etc.) informed by insights from the cluster modeling. A graphical representation of the decision tree model was produced.

Findings

Clustering Results

The clustering of the imputed data resulted in the dendrogram (Figure 1). Given the number of observations, the bottom of the dendrogram is undecipherable. However, it exhibits several possible cut points for producing clusters. I chose to cut the dendrogram at four clusters, as the next best options were two or eight. I decided that two clusters would not provide enough granularity to be meaningful, while eight clusters would be too complex to interpret politically.

Figure 1. Hierarchical Clustering Output

I produced cluster means at the 4-cluster cut and selected the columns that were most relevant to my original research goal of understanding the Belarusian polity in relation to Russian influence and informing political messaging strategies.

The results are exhibited in table 1. 

Table 1: Cluster Means

Please note that all cluster interpretations are in comparative, not absolute, terms. For example, when the analysis states “X cluster views themselves as ethnically Russian at a high rate”, this should be interpreted in comparative terms - “X cluster reports high identification with the Russian ethnicity compared to other clusters produced by the clustering algorithm.”

Cluster 1:

With an average age of 55.33 years, this is our oldest cluster and demonstrates a mixed attitude towards political activism. While they express moderate willingness to participate in mass protests (4.5%) and contact public organizations (7.5%), their engagement in voting is the highest among the clusters (18.5%) when asked how they would address dissatisfaction with the government. This cluster reported that the lack of democracy and freedom of speech was a bigger problem in Belarus than other clusters (24.5%) and similarly supported the release of political prisoners (21.6%). They were most likely to report that Belarus was in need of reform (67.5%), considered themselves ethnic Russians (27%), and supported Belarus and Russia forming a union state with a single currency, president, and parliament (40.2%). They also expressed moderate support for Ukraine in the war (19.3%) and a preference for increasing state authority more than other clusters (18.8%).

Cluster 2:

With an average age of 24.5 years, this cluster presents a younger demographic. Their engagement with public organizations and voting is comparatively lower than all other clusters (9.6%). This cluster displays a notable preference for EU membership (42.8%) over a Belarus-Russia union state (27.5%), along with the lowest average agreement that Belarus needs reform (27.8%). This youth cluster expressed the least support for Ukraine in the war among all clusters (14.5%) and the least support for the release of political prisoners as an important priority for Belarus (7%). They were most supportive of increasing army funding (26.1%) and police funding (27.3%) and least likely to believe that Russia is more corrupt than Belarus (28.7%).

Cluster 3:
With an average age of 32.8 years, this cluster demonstrates a moderate inclination to contact public organizations (8.6%) and express grievances through voting in elections (11.8%). Their willingness to participate in mass protests (2.6%) is the lowest among all clusters. They were the second least likely to consider themselves ethnically Russian (9.1%) and least supportive of the idea of Russia and Belarus forming a single union (26.6%). They preferred Belarus becoming an EU member state (40.8%) over forming a union with Russia. This cluster was most supportive of Ukraine in the war (24%) and least supportive of increasing army funding (11.2%) among all clusters.

Cluster 4:

With an average age of 42.4 years, this represents the second oldest demographic among the clusters. They demonstrate the highest willingness to participate in mass protests (9.6%). This cluster was more supportive of Belarus becoming an EU member state (40.5%) than forming a union with Russia (29.4%), though notably more open to the latter compared to clusters 2, 3, and 4, second only to the oldest cluster. They were least likely to consider themselves ethnically Russian (6.1%) and least likely to believe that there should be reforms to increase the authority of the state in Belarus (10.6%). This cluster also agreed the most with the idea that Russia is more corrupt than Belarus.

Tree-Based Modeling Results

The first run of my tree model on the test data resulted in a tree with nine terminal nodes, a residual mean deviance of 0.803, and a misclassification error rate of 0.17. A value of one for the variable being predicted indicates that the respondent opposes Belarusian integration into the European Union. The tree reveals several politically intuitive results. Respondents classified into terminal nodes with the prediction "1," indicating opposition to EU integration, were less supportive of Ukraine in the Russia-Ukraine war, more open to moving to Russia, older than 29, and did not believe Belarus needed reform or that democracy and freedom of speech were problems in Belarus. Another group opposed to EU integration included respondents who were less supportive of Ukraine, more open to moving to Russia, older than 29, did not think Belarus needed reform, and were not open to residing anywhere other than Belarus. Intuitively, no one who was highly supportive of Ukraine in the Russia-Ukraine war was classified into the anti-EU integration terminal nodes. Additionally, people who were minimally supportive of Ukraine, more open to moving to Russia, and younger than 29 were also against EU integration.

Some interesting insights from this tree include frequent splits based on the age variable and variables regarding potential residency, indicating these are important indicators of openness to EU integration.

Conclusion 

This research sought to answer the following: What are the attributes of a promising target audience in Belarus for pro-democratization political messaging? A general takeaway from this research is that statistical learning methods used on open-source survey datasets can offer valuable insights for political messaging strategies. Depending on the angle and goals of the political messaging in question, several insights gleaned from this research could inform targeting strategy with specificity.

Using opposition to European integration as a proxy for general opposition to political alignment with the West, the tree model indicates that people who support Ukraine in the Russia-Ukraine war are far more likely to be open to European integration. On the other hand, among those less supportive of Ukraine, some subsets may still be open to Western politics: people under the age of 30 or those who report they would not move to Russia, for example.

The clustering provides further guidance on the attributes of a target Belarusian audience for political messaging, with Cluster 4 standing out as a promising candidate. Cluster 4 demonstrates the highest willingness to participate in mass protests among all clusters, indicating strong discontent with the current political situation and a readiness to take to the streets for change. This cluster displays a higher preference for Belarus becoming an EU member state, suggesting alignment with Western values and a desire for closer integration with Europe. Unlike other clusters, Cluster 4 is least likely to support reforms that increase the authority of the state (10.6%), indicating skepticism towards centralized power and authoritarian tendencies, which could make them more receptive to messages advocating for democratic reforms or alignment with the West.

Works Cited

Eighteenth survey wave. (n.d.). Retrieved March 29, 2024, from https://en.belaruspolls.org/wave-18

Mankoff, J. (2024, April 2). Next Door to Ukraine, Moscow’s Grip Is Tightening. Foreign Policy. https://foreignpolicy.com/2023/08/22/russia-influence-belarus-georgia-moldova-putin-ukraine-war-empire/

U.S. Army Information Operations: Concept and Execution. (n.d.). Retrieved May 5, 2024, from https://irp.fas.org/agency/army/mipb/1997-1/mcconvl.htm