Introduction

The aim of this work is to analyze demographic and social factors differentiating the voters of the two largest Polish political parties - Law and Justice (PiS) and Civic Coalition (KO). Knowing the specificity of these two parties, the selected variables should allow for the verification of their policies. Beyond mere differentiation, this analysis aims to uncover the core factors that set these parties apart and illuminate the specific social demographics drawn to their political approaches. For the parties, it may also be an indication of what they can focus on in order to gain greater electoral support.

For this purpose, various unsupervised learning algorithms were used in this study. The first focus was on social factors, performing clustering using K-means and PAM and dimension reduction using MDS and PCA. By initially concentrating on social factors, the study engages in clustering and dimension reduction techniques, extending its examination to Polish regions. This analysis seeks to discern the root causes that contribute to their unique political landscapes.

Dataset description

This database was built by the author on the basis of data from the Polish Central Statistical Office of Government Administration - GUS and from the official source of 2023 election results - GOV.

The database consists of 16 observations and 15 variables.The observation units are the regions of Poland’s administrative division. In addition to the voting results for the two most important political parties, other variables include various demographic factors that influence citizens’ voting decisions. Apart from data about 2023 election, other variables are from year 2022 (last available on GUS).

Variables

The data set consists of the following variables:

  • Population density - population per 1 kilometer;
  • Unemployment rate - percentage of people in the labour force who are unemployed;
  • Life expectancy(men) - number of years a person can expect to live (among men);
  • Beds in hospitals - beds in hospitals (per 10,000 people);
  • Nurseries - percentage of children cared for in nurseries
  • Natural increase rate - difference between the birth rate and the death rate (per 1,000 people)
  • Disposable income - amount of money that an individual has to spend or save after income taxes in relation to the national average (Poland = 100)(percentage)
  • Social benefits - people who were granted social assistance benefits (per 10,000 people)
  • Gross salary - average monthly gross wages in relation to the national average (Poland = 100)(percentage)
  • Pensioners - percentage of retirees and pensioners in the population
  • Urbanization - percentage of population living in urban areas
  • Founds form EU - value of contracts for co-financing from EU funds per capita (Poland = 100)
  • Voter turnout - participation rate of Poland’s 2023 election
  • Voting for PiS - percentage of votes cast for party “Law and justice”
  • Voting for KO - percentage of votes cast for party “Citizen Coalition”

Libraries

library(clusterSim) 
library(factoextra)
library(FeatureImpCluster)
library(flexclust)
library(dendextend)
library(psych)
library(smacof)
library(gridExtra)
library(hopkins)
library(sf)
library(ggplot2)
library(dplyr)
#devtools::install_github("o1iv3r/FeatureImpCluster")

Dataset preparation

DATA <- read.csv("Poland.csv", header=TRUE, sep=",", dec=".")
rownames(DATA)<-DATA[,2]
DATA <- DATA[,c(-1,-2,-3,-19,-20)]

#normalization of data
DATA.new <- data.Normalization(DATA, type="n1", normalization="column")

Analysis of social factors

Hopkins statistics

Veryfing how well the data can be clustered, can be conducted using Hopkin’s statistic. Null hypothesis: dataset does not contain meaningful clusters. Alternative hypothesis: dataset constains meaningful clusters.
Calculated values 0-0.3 indicate regularly-spaced data. Values around 0.5 indicate random data. Values 0.7-1 indicate clustered data.

hopkins(t(DATA.new), m=nrow(DATA.new)/10)
## [1] 0.9941266

The value is close to 1, which means that data is clusterable.

Using another function we will verify clustering tendency. We again have Hopkin’s statistic - if the value of is close to 1 (far above 0.5), then we can conclude that the dataset is significantly clusterable.

get_clust_tendency(t(DATA.new), 10, graph=TRUE, gradient=list(low="brown1", mid="snow", high="royalblue3"), seed = 123)
## $hopkins_stat
## [1] 0.5832039
## 
## $plot

These values indicate the possibility of clustering but not very significant. On the plot we see two clearer patterns.

Optimal number of clusters

fviz_nbclust(t(DATA.new), FUNcluster=kmeans, method="silhouette")
fviz_nbclust(t(DATA.new), FUNcluster=kmeans, method="gap_stat")

We will only do 2 clusters, but in this case it may be even more useful because we want to analyze the two most influential political parties.

K-means and PAM clustering for social factors

km<-eclust(t(DATA.new), "kmeans", hc_metric="euclidean", k=2)
pam<-eclust(t(DATA.new), "pam", hc_metric="euclidean", k=2) 
fviz_cluster(km, main="K-means / Euclidean", xlim=c(-5, 5), ylim=c(-5, 5), labelsize=10, repel=TRUE)
fviz_cluster(pam, main="PAM / Euclidean", xlim=c(-5, 5), ylim=c(-5, 5), labelsize=10, repel=TRUE)

What is the most important about these results is that we see two very different clusters, each of which includes one political party. These clusters define the attributes of their electorates.

The results are in line with intuition, as Party PiS adheres to traditional views, and targets its policies to the elderly and those of lower socioeconomic status. It is apparent that the party’s group has a higher share of pensioners. PiS introduced a social benefit program, and the people to whom the benefits went also seem more likely to be the party’s voters in 2023. Another aspect is the unemployment rate - in regions where it is higher, more people vote for PiS. This may be due to the welfare state policies that the party pursue, as well as greater support in villages and smaller urban centers, where the unemployment rate may be higher.

People from regions with higher urbanization rates and higher population densities, as well as those with higher economic status, are more likely to vote for the Civic Coalition party. What is interesting is that the goal of Party Law and Justice’s main social benefit program was to increase the birth rate, and in regions where it is higher, people were more likely to vote for KO.

It is also interesting to note that voter turnout was higher in regions more likely to vote for KO. This may confirm that the motivation to change power in these particular regions influenced the outcome of the 2023 election and the removal of the ruling party from power.

We can examine the similarity between clustering using K-means and PAM.

Rand Index and Jaccard similarity - comparing K-means and PAM

km<-eclust(t(DATA.new), "kmeans", hc_metric="euclidean",k=2)
pam<-eclust(t(DATA.new), "pam", hc_metric="euclidean",k=2)
comPart(km$cluster, pam$cluster) 
## ARI  RI   J  FM 
##   1   1   1   1

Overlap is 100% which means that both K-means and PAM give the same results of clustering.

Now let’s see how well the data is clustered using Silhouette plot.

fviz_silhouette(km)
##   cluster size ave.sil.width
## 1       1    9          0.25
## 2       2    6          0.30
fviz_silhouette(pam)
##   cluster size ave.sil.width
## 1       1    9          0.25
## 2       2    6          0.30

There are no negative silhouette values, which means the clustering is correct.
Additionally we can examine which regions are the most important for clustering.

Variable importance for K-mans

km<-kcca(t(DATA.new), k=2)
FeatImpKmeans<-FeatureImpCluster(km, as.data.table(t(DATA.new)))
plot(FeatImpKmeans)

Śląsk has a substantial impact on the formation of clusters within this dataset. Let’s analyse the dispersion within clusters.

d1<-cclust(DATA.new, 2, dist="euclidean")
stripes(d1) 

It is hard to decide which cluster is more dispersed, which could give the information about whose electorate is more diverse.

Hierarchical clustering for social factors

dend.f<-hcut(t(DATA.new), k=5, stand=FALSE)
dend.f$cluster
##         Population.density          Unemployment.rate 
##                          1                          2 
##       Life.expectancy.men.          Beds.in.hospitals 
##                          3                          4 
##                  Nurseries           Natural.increase 
##                          5                          3 
##          Disposable.income            Social.benefits 
##                          1                          2 
##               Gross.salary                 Pensioners 
##                          1                          4 
##               Urbanization             Founds.from.EU 
##                          5                          2 
##              Voter.turnout Voting.for.Law.and.Justice 
##                          1                          4 
## Voting.for.Civic.Coalition 
##                          5
fviz_dend(dend.f, rect=TRUE, horiz=TRUE, labels_track_height=10)

This dendrogram give us a clearer view of the clustering. We can again conclude that the electorate of KO is most influenced by the percentage of population living in urban areas and percentage of children cared for in nurseries. Whereas electorate of PiS is most influenced by the percentage of retirees and pensioners in the population and access to health care (expressed in the availability of hospital beds).

fviz_silhouette(dend.f)
##   cluster size ave.sil.width
## 1       1    4          0.22
## 2       2    3          0.36
## 3       3    2          0.53
## 4       4    3          0.19
## 5       5    3          0.20
fviz_cluster(dend.f)

Even though this time instead of 2 clusters we have 5, there is no negative value of silhouette.

MDS - Multidimensional scaling for social factors

dist1<-dist(t(DATA.new))
fit.data<-mds(dist1, ndim=2,  type="ratio") 
fit.data
## 
## Call:
## mds(delta = dist1, ndim = 2, type = "ratio")
## 
## Model: Symmetric SMACOF 
## Number of objects: 15 
## Stress-1 value: 0.224 
## Number of iterations: 55
plot(fit.data, plot.type = "stressplot")

Stress point Stress points give the information about goodness-of-fit in MDS, usually expressed as a percentage. It ranges from 0% to 100%, where lower values indicate a better fit between the observed dissimilarities and the distances in the MDS solution.

Values: 0.20 = poor, 0.10 = fair, 0.05 = good, 0.025 = excellent, 0.00 = perfect. In this data set most variables are between 10 and 4 % which indicates rather good quality of the model. There are no clear outliers. Stress-1 value close = 0.224 suggests a relatively decent fit for the MDS.

par(mar=c(5,5,4,2))
plot(fit.data, pch=21, cex=as.numeric(fit.data$spp), bg="lightsteelblue", main="MDS for selected variables", xlim=c(-1.2, 1.2), ylim=c(-0.8, 0.8))

With the help of the MDS, we can better see what factors are more correlated with each other and what factors are far from each other. We still see the division into two main clusters. An interesting observation more visible on this perceptual map is that voting for Citizen Coalition is highly correlated with the level of urbanization and access to nurseries.

This data and chart also confirm the trend that people with higher economic status are more likely to participate in elections. We can deduce this by the fact that “Voter turnout” indicator is very close to “Gross salary” on MDS perceptual map.

PCA - principal component analysis for social factors

pca.f<-principal(DATA.new, nfactors=4, rotate="varimax")
print(loadings(pca.f), digits=3, cutoff=0.4, sort=TRUE)
## 
## Loadings:
##                            RC3    RC2    RC1    RC4   
## Nurseries                  -0.637         0.426       
## Urbanization               -0.693                0.575
## Voting.for.Law.and.Justice  0.940                     
## Voting.for.Civic.Coalition -0.951                     
## Life.expectancy.men.        0.490 -0.741              
## Beds.in.hospitals                  0.834              
## Natural.increase                  -0.913              
## Pensioners                         0.917              
## Population.density                        0.687  0.412
## Unemployment.rate           0.463        -0.684       
## Social.benefits                          -0.895       
## Founds.from.EU                           -0.824       
## Disposable.income                                0.883
## Gross.salary                              0.476  0.728
## Voter.turnout                             0.456  0.747
## 
##                  RC3   RC2   RC1   RC4
## SS loadings    3.412 3.281 3.165 2.695
## Proportion Var 0.227 0.219 0.211 0.180
## Cumulative Var 0.227 0.446 0.657 0.837

SS loadings - indicate the strength and direction of the relationship between the original variables and the principal components. Larger SS loadings indicate that the principal component captures more variance in the original data. We can see that RC3 captures the most variance.

Cumulative variance - gives a sense of how much of the total variance is explained as we include more principal components. In this case, the first two components (RC3 and RC2) together explain 45% of the variance, which seems a good score. Total variance explained is 83%.

pca2<-princomp(DATA.new)
fviz_pca_var(pca2, col.var="lightcoral", repel=TRUE, ylim=c(-1,1), xlim=c(-1,1))

On this perceptual map positively correlated variables are grouped together. Negatively correlated variables are positioned on opposite sides of the plot origin. We can see that there is a contrary relation between Voting for PiS and Urbanization rate or nurseries. Other connections confirm that this method works well - for example Gross salary and Unemployment rate, Natural increase and population of Pensioners.. Interestingly there is a contrary relation between Voter turnout and Unemployment rate or people granted with Social benefits.

fviz_eig(pca2)

We can see the percentage of explained variance on y-axis of this plot.

var<-get_pca_var(pca2)
a<-fviz_contrib(pca2, "var", axes=1, xtickslab.rt=90) # default angle=45°
b<-fviz_contrib(pca2, "var", axes=2, xtickslab.rt=90)
gridExtra::grid.arrange(a,b,top='Contribution to the first two Principal Components')

Here we can verify which variables contribute the most to our principal components. For dimension 1 - unemployment rate, gross salary and voter turnout. For dimension 2 - life expectancy, natural increase and percentage of pensioners.

Analysis of regions

In this part, the same analysis was carried out, but this time the goal was to examine the differences and similarities between the regions of Poland.

Hopkin’s statistics

hopkins(DATA.new, m=nrow(DATA.new)/10)
## [1] 0.9997634
get_clust_tendency(DATA.new, 10, graph=TRUE, gradient=list(low="brown1", mid="snow", high="royalblue3"), seed = 123)
## $hopkins_stat
## [1] 0.5705516
## 
## $plot

Again there is a possibility of clustering but not very significant. On the plot we don’t see clear patterns.

Optimal number of clusters

fviz_nbclust(DATA.new, FUNcluster=kmeans, method="silhouette")
fviz_nbclust(DATA.new, FUNcluster=kmeans, method="gap_stat")

I decided to make 3 clusters instead of four. This decision was promped by the occurrence of a negative silhouette value in 4 clusters.

K-means clustering for regions

km<-eclust(DATA.new, "kmeans", hc_metric="euclidean", k=3)
fviz_cluster(km, main="K-means / Euclidean", xlim=c(-5, 5), ylim=c(-5, 5), labelsize=10, repel=TRUE)
fviz_silhouette(km)
##   cluster size ave.sil.width
## 1       1    4          0.23
## 2       2    6          0.15
## 3       3    6          0.24

For regions analysis I choose to do only K-means clustering. In 3 clusters there are no negative sillhouette values, so the clustering seems well prepared.

Hierarchical clustering for regions

dend.f<-hcut(DATA.new, k=3, stand=FALSE) 
dend.f$cluster
##        DOLNOŚLĄSKIE  KUJAWSKO-POMORSKIE             ŁÓDZKIE           LUBELSKIE 
##                   1                   2                   1                   2 
##            LUBUSKIE         MAŁOPOLSKIE         MAZOWIECKIE            OPOLSKIE 
##                   1                   3                   3                   1 
##        PODKARPACKIE           PODLASKIE           POMORSKIE             ŚLĄSKIE 
##                   2                   2                   3                   1 
##      ŚWIĘTOKRZYSKIE WARMIŃSKO-MAZURSKIE       WIELKOPOLSKIE  ZACHODNIOPOMORSKIE 
##                   2                   2                   3                   1
fviz_dend(dend.f, rect=TRUE, horiz=TRUE, labels_track_height=10, k_colors = c("dodgerblue3", "indianred3", "palegreen3"),  # Change line colors
  color_labels_by_k = TRUE)

Knowing the specifics of Polish regions, this clustering result seems interesting and consistent with the frequent historically determined division. Red cluster points to cities that are the largest urban centers in terms of economy, industry and the scientific world. Blue cluster presents the regions of Poland’s Eastern Bloc, which were under Russian occupation for a long time in history. On the other hand, the green cluster presents the regions of Poland’s Western Bloc, which was under Prussian occupation. These regions differ in terms of types of their economy and agriculture, which, as this analysis shows, also differentiates these regions in terms of demographic and social factors. This, in turn, means that historical conditions still have a major impact on Polish societies and their distinctive cultural and economic identity.

DATA2 <- read.csv("Poland.csv", header=TRUE, sep=",", dec=".")
DATA2 <- DATA2[,c(2,19,20)]

map_poland <- st_read("/Users/emiliaselwa/Desktop/PL/pol_adm_gov_v02_20220414_SHP/pol_admbnda_adm1_gov_v02_20220414.shp")
map_poland$ADM1_PL <- toupper(map_poland$ADM1_PL)

merged_data <- merge(map_poland, DATA2, by.x = "ADM1_PL", by.y = "Region")

ggplot() +
  geom_sf(data = merged_data, aes(fill = Color), color = "white", size = 0.2) +
  scale_fill_identity() +
  theme_minimal()

The map above presents described relationships and division of regions according to demographic and social factors. It can be noticed that the agglomeration centers marked in red, separate two parts of the country which have different demographic characteristics.

Variable importance for K-mans

km.f<-kcca(DATA.new, k=3)
FeatImpKmeans<-FeatureImpCluster(km.f, as.data.table(DATA.new))
plot(FeatImpKmeans)

This plot enables to recognize which variables had the biggest impact on the formation of clusters of regions. We see that voter turnout and voting for both PiS and KO, were important factors for the analysis. Urbanization, unemployment rate, and disposable income, also had a significant influence for the clusters.

MDS - Multidimensional scaling for regions

dist1<-dist(DATA.new)
fit.data<-mds(dist1, ndim=2,  type="ratio") 

fit.data
## 
## Call:
## mds(delta = dist1, ndim = 2, type = "ratio")
## 
## Model: Symmetric SMACOF 
## Number of objects: 16 
## Stress-1 value: 0.208 
## Number of iterations: 47
plot(fit.data, plot.type = "stressplot") 

Most regions vary between 10 and 4% of stress proportion, which indicates rather good quality of the model. Two regions - Opolskie and Śląskie seem to apper as outliers. Stress-1 value = 0.208 suggests a relatively decent fit for the MDS.

par(mar=c(5,5,4,2))
plot(fit.data, pch=21, cex=as.numeric(fit.data$spp), bg="lightsteelblue", main="MDS for regions", xlim=c(-1.2, 1.2), ylim=c(-0.8, 0.8))

The significance of Śląskie region in variable clustering suggests its unique importance in the broader regional context. On this perceptual map we see that Śląskie is situated visibly further than other regions. This region has been historically associated with heavy industry, coal mining, and metallurgy. The industrialization of the region may have shaped its economic landscape differently compared to other parts of Poland. Economic activities and the presence of specific industries can have a profound impact on the demographic composition and social dynamics of a region.

Conclusions

This study focused on analyzing demographic and social factors that differentiate voters of the two major Polish political parties - Law and Justice (PiS) and Civic Coalition (KO). Utilizing unsupervised learning algorithms such as K-means, PAM, MDS, and PCA, the research identified distinct clusters corresponding to each political party.

The findings align with expectations, revealing that PiS attracts predominantly elderly individuals and those with lower economic status. The party’s social benefit program, targeted at this demographic, seems to have influenced electoral support. On the other hand, KO garners support from regions characterized by higher urbanization, population density, and economic status. Notably, the higher voter turnout in regions favoring Civic Coalition suggests a motivation for change, influencing the outcome of the 2023 election and the removal of the ruling party from power. These results provide valuable insights for both political parties to understand their electorate and tailor their strategies for greater electoral success.

The clustering results revealed intriguing patterns that align with historical divisions in the country. Historical contexts contribute to variations in economic types, agriculture, and, consequently, demographic and social factors. The analysis underscores the enduring impact of historical conditions on Polish society. Notably, voter turnout, support for PiS and KO, urbanization, unemployment rate, and disposable income emerged as crucial factors shaping the clusters.