Comparison of turnout in Italian parliamentary elections: a two-stage affinity propagation approach

ABSTRACT

This paper presents a methodology for temporal contextual analysis of electoral outcomes in the presence of redistricting. Affinity propagation is used to aggregate electoral districts, thus deriving spatial units comparable over electoral waves. The methodology is based on travel-to-work flows and socio-economic aggregated data. Comparability of clusters enables analyses of electoral trends. A map of electoral turnout variation in Italian parliamentary elections held in 2018 and 2022 illustrates the proposed approach.

KEYWORDS:

• Electoral boundaries redistricting
• temporal contextual analysis
• travel-to-work
• affinity propagation

1. Introduction

Voter turnout is decreasing across Europe as well as in most democratic countries, OECD (2019). National parliamentary elections in Italy confirm this trend. While in 2006, about 83 per cent of registered persons cast a vote, this percentage steadily decreased by 3.5 points on average at each electoral round, see Supplementary Table 1. In the 2022 elections, the highest breakdown in the Italian Republic’s history occurred. Indeed, turnout decreased by more than 9 percentage points with respect to the previous elections. Large differences in electoral turnout raise concerns for the legitimacy of democratic government and for societal representation, Boyle (2024). Understanding the reasons behind absenteeism¹ is one of the core objectives of electoral scientific studies that focus on socio-economic conditions and trust towards political systems. Italian NUTS2² Regions also exhibit different turnout levels and trends. Supplementary Map 1 shows the voter turnout differences between 2018 and 2022 elections. The largest variation with respect to the previous parliamentary elections is found in the Molise region (−15.1 per cent); the smallest variation is registered by the Emilia-Romagna region (−6.3 per cent). The North–South gradient represented in Supplementary Map 1 supports the idea that interactions between socio-economic conditions and spatial population distribution affect voting behaviour.³ Indeed, lower turnout levels may be associated with poorer areas, lower social capital and institutional inefficiency, Istat (2023). Anyway, the North–South gradient does not completely explain the national turnout drop. Moreover, a finer analysis of spatial voting preferences needs more territorial detail, at least at the electoral district level.

An important factor shaping electoral participation is the political context. Both party system and electoral rules contribute, through different mechanisms, to people’s decision to cast ballots. In Italy, in the last 30 years, the national parliamentary electoral system suffered many changes. The frequency of changes might be a cause of declining voter turnout since it undermines trust in political parties and increases information costs.

The current Italian electoral system is a mixed one, including single-member first-past-the-post and multi-member proportional districts. This holds for both legislative bodies, i.e. Chamber of Deputies and the Senate. In Italy, the boundaries of electoral districts are drawn by an extra-legislative delegation called the Electoral Districts Board. The electoral boundaries are enforced 30 days before coming into force of the Law, upon parliamentary opinion. The Italian Electoral Law 165/2017⁴ defines two sets of rules for designing district boundaries. The mandatory rules regard equal population criteria, compactness and coherence with territorial administrative units. The law also states that electoral districts should be coherent with other functional areas (e.g. labour market areas) and should exhibit socio-economic homogeneity. It should be stressed that the Italian districts could not be drawn based on any information related to the partisan system, political choice or participation. On the contrary, the Italian electoral law emphasises communities-of-interest standards, see also Winburn and Wagner (2010). Geographical databases, the most recent population Census and socio-economic indicators stemming from official statistics are used for designing electoral boundaries, Electoral Districts Board (2017). While researchers sometimes ascertained a positive impact of an extra-legislative body delegated to boundaries definition on electoral participation, e.g. Carson et al. (2014), the Italian national parliamentary vote turnout trend does not support this hypothesis.

In 2020, the Italian electoral law heavily diminished the total assembly size while preserving the other electoral rules. For example, the House of Representatives counts 400 deputies with respect to the previous 630 members. This led to a reduction in the number of electoral districts. The number of single-member first-past-the-post districts is 147 w.r.t. the previous 232 ones. As the electoral districts are functional geographies, their boundaries were redesigned in order to take into account the new assembly size. Supplementary Map 2 depicts the 2018 and 2022 electoral districts. In terms of population, the current districts are 1.5 times larger than the 2018 districts. Descriptive statistics are shown in Supplementary Table 2. Given the legal constraints and population demography, newer electoral districts are not necessarily drawn by dissolving older districts. Supplementary Map 3 shows an example of 2018 district that was entirely dissolved into a new electoral district and an example of a district that was distributed over different newer districts.

Researchers largely discuss the effect of redistricting on electoral participation. Turnout and redistricting are linked by mechanisms related to competitiveness, incumbency or party environment. Firstly, changes in districts’ boundaries cause changes in electoral closeness, i.e. competitiveness. According to rational choice theories of voting (Downs, 1957), individuals are more likely to turnout when they believe their vote is decisive. In other words, higher turnout levels are registered in the presence of competitive elections. Several studies questioned the strength of the turnout–competiveness relationship, see, for example, Blais (2000) and Geys (2006); instead, other researchers enhanced the theory by using suitable measurements, Simonovits (2012).

Incumbency is a second mechanism connecting redistricting and turnout. When citizens change their electoral district, they might be subject to increased information costs in order to get more familiar with new candidates; they might be unable to evaluate their governmental performances, etc. Consequently, they might be prone to absenteeism, Hunt (2018) and McKee (2013). Moreover, in the presence of redistricting, the party environment might completely change. Party composition of a district and competiveness are obviously related. Indeed, voters affiliated with majoritarian parties in an old district might be redistricted into districts where their party is in the minority. Their votes are thus diluted by votes cast outside their old community of interest, see Born (2022) and Scarrow (1999).

In the literature on geographical influences of voting, contextual approaches, i.e. those claiming the existence of a neighbourhood effect, play an important role. The contextual perspective explores and accounts for the relationships between political behaviour and place and location. Indeed, the political attitudes of individuals may be influenced by the context within which ‘individuals live out their daily lives’ (Jones et al., 1992, p. 346): the family, the sector in which they work, their class position, their place of residence, etc. Starting from the work of Agnew (1996) and Pattie and Johnston (2000) who introduced the ‘speak together, vote together’ approach, the theoretical and empirical literature explored the role of human interactions in shaping political behaviour. Several studies show that the neighbourhood effect could determine increased political homogeneity through person-to-person interactions (families, friends and acquaintance circle) or organisationally based interactions (workplaces, churches, etc.), Flint et al. (2000), Johnston et al. (2016), Munis (2021), Vilalta y Perdomo (2004), to cite a few. Human interactions together with the spatial context within which interactions occur characterise the neighbourhood effect, Jones et al. (1992).

Currently, at least in Italy, limited comparative data allowing contextual analysis is available. Indeed, the frequency of redesigning of electoral districts’ boundaries limits their temporal coverage. As districts change, turnout trends at the district level may not be analysed.

This paper proposes a socio-economic classification of the Italian districts by means of a data-driven clustering algorithm. The comparison of outcomes of different electoral rounds is enabled within each class. Researchers could analyse turnout trends among similar districts, i.e. by classes of districts. In this work, statistical indicators stemming from official statistics allow the delineation of the neighbourhood (spatial context) by considering demographic-cultural, social-economic and commuting to work dimensions.

The paper maps and analyses the classification of the single-member first-past-the-post Italian electoral districts. Data from the 2018 and 2022 elections illustrate the classification potential.

The next section presents the data and clustering steps. The Results and Discussion section presents the Main Map with its interpretation. The last section, Conclusions, summarises the paper and traces future research directions. The output of the paper is the Main Map, which visualises the spatial distribution of the electoral turnout variation on aggregated electoral districts.

2. Data and methods

2.1. Data

This paper does not use any individual data. The contribution is based on the single-member electoral districts of the Italian 2018 and 2022 parliamentary elections. The district boundaries are available at Istat’s website www.istat.it.

Personal and organisational interactions were modelled through travel-to-work commuting flows at the district level; these flows were aggregated from the 2011⁵ Origin–Destination matrix at the census-tract level. Indeed, independently of the electoral wave, each Italian census-tract is included in a unique district. Consequently, by means of spatial overlaying, the commuting flows were simply summed at the district level, for each election wave. The census tracts were further disregarded in our research.

The spatial context was modelled using demographic, cultural, social and economic characteristics affecting voting decisions, see Lijphart (1997) and Kulachai et al. (2023). As scientific literature points out, individuals with higher levels of education, older citizens or wealthier persons tend to vote more, Smets and Ham (2013). From a utility point of view, rich persons might influence more policymakers’ decisions; thus, they might be more interested in voting. Due to the efforts needed to collect and analyse (cognitive skills) information, voting might be more onerous than absenteeism for poor or less educated individuals. As for older citizens, when compared to younger ones, they are more familiar with voting also grace to their higher level of civic education. The indicators employed in the analysis, Supplementary Table 3, are available on Istat’s website. All attributes are continuous variables. It is important to underline that these socio-economic indicators are exactly those used in Electoral Districts Board (2017) and catch different phenomena. They are neither correlated nor overlapping. The three measures of education, for example, refer to different subpopulations: individuals aged 18–24 that achieve only secondary education; illiterate individuals aged 15 or more and individuals aged 19 or more holding tertiary education qualification.

Electoral data was downloaded from the Italian Interior Minister website, https://www.interno.gov.it/it. In order to describe electoral competiveness, incumbency and party environment, the following indicators were calculated:

• voter turnout,

• difference between the first two parties,

• number of parties exceeding three percent of votes,

• percentage of male candidates and

• percentage of candidates born in each district.

2.2. Methods

The proposed methodology achieves comparable super-districts by subsequently applying affinity propagation (AP) on two types of data. AP is a modern unsupervised deterministic clustering method that groups data points together based on exemplars, representative data for each group, see Frey and Dueck (2007). AP is an iterative algorithm based on a similarity matrix between data points. Initially, each data point is set to be an exemplar. Using the similarity matrix, the iterations update two message-passing quantities called ‘responsibilities’ and ‘availabilities’. Responsibilities indicate how evident point k is to be an exemplar for point i, accounting for other candidate exemplars for point i. Availabilities measure the opportunity to select point k as an exemplar for point i, considering the support from other points. At each iteration, for each i, its exemplar is identified as the data point maximising the sum of availability and responsibility. The data points sharing the same exemplar are assigned to the same cluster. AP repeatedly identifies exemplars and clusters until convergence, see Frey and Dueck (2007) for more details. With respect to other clustering methods, AP overcomes issues with thresholds. Indeed, it has the advantage that the number of clusters is not a-priori defined. Furthermore, the exemplar, an actual datum point, is representative of the entire cluster.

Firstly, AP is applied to travel-to-work flows in order to aggregate electoral districts. Commuting naturally creates personal and organisational neighbourhoods and it favours informal information channels, through which the ‘speak together, vote together’ paradigm takes place, Agnew (1996), Kim et al. (2003), and Wicki et al. (2019). These information channels are likely to be more efficient when spatial units are closer. By construction, original districts are contiguous; at the same time, AP does not necessarily maintain spatial contiguity. Consequently, in the first stage, AP is properly applied to commuting between electoral districts. For each electoral wave, the first-stage AP uses the intensity of travel-to-work flows as the distance between districts. From the viewpoint of contextual analysis, the body of information transmitted (on candidates, governmental efficiency, etc.) increases with the number of persons travelling between districts (thus augmenting the probability of influencing electoral behaviour). It may be stated that first-stage clusters (of electoral districts) delimit the geographic scope of voters’ interactions. Aggregating districts based on travel-to-work flows also ensures that there are limited external influences. Indeed, if individuals live and work within self-contained areas, there are less daily life chances to speak to persons or interact with institutions outside that area.

The second stage of AP concerns the socio-economic dimension, see Heumann et al. (2020) for an analogous approach. Indeed, only socio-economic homogeneous spatial units would allow robust temporal contextual comparisons. For each electoral wave, the clusters obtained in the first step are submitted to AP using the indicators in Supplemenatry Table 3. The standard negative squared distance between clusters is used as similarity measure, Frey and Dueck (2007). These indicators relate to many factors significantly affecting electoral behaviour, i.e. education, age, employment, migration, etc.

For each electoral wave, the two-stage AP identifies a set of super-districts, i.e. groups of electoral districts. By spatially overlaying the two sets of super-districts, their similarity may be assessed. If necessary, AP may be refined in order to improve the resulting spatial similarity. For example, users might control for the number of clusters, Frey and Dueck (2007).

Once comparable super-districts are obtained, researchers may assess between wave variations of different electoral indicators. In the absence of comparable spatial units, temporal variations cannot be analysed.

Using only aggregated data at the district level, the proposed methodology may be applied in many settings, i.e. local or parliamentary elections, in many countries, etc. Furthermore, the methodology avoids measurement errors characterising more often individual-level data, e.g. sampling or systematic errors. Finally, the methodology does not use any political information as input data. Consequently, researchers using such super-districts should not be worried about hidden relationships between spatial units and phenomena under study. Comparable super-districts are neighbourhoods where voters are as homogeneous as possible with respect to their individual characteristics and commuting behaviour.

In this paper, electoral data are used only to validate the electoral outcomes at the super-district level. Average income levels are used for the same purposes.

3. Results and discussion

The Main Map shows the super-districts obtained when using the above methodology on 2018 and 2022 Italian electoral districts. Supplementary Table 4 shows the number of clusters obtained in each AP stage. Italian data and the standard AP⁶ setting generate seven clusters for both sets of districts. By spatially overlaying 2018 and 2022 super-districts, their spatial extensions show an extremely significant similarity. Consequently, the Main Map depicts a single partition. Depending on the case study, the super-districts might be further processed. For example, by splitting or dissolving operations, researchers could constrain super-districts to contiguity.

This section describes the clusters by means of socio-economic and electoral data, see Supplementary Table 5. The Main Map shows the turnout variation in Italian parliamentary elections, 2018–2022. The map was obtained with QGIS version 3.22.13 (coordinates Reference system WGS84 /UTM zone 32N, EPSG:32632).

A special status NUTS Region, Trentino-Alto Adige, coincides with the cluster SD1. While employment is the main driver, high education levels, high income and low population density further feature this cluster. As these socio-demographic characteristics affect turnout, this Region shows one of the smallest absenteeism values and variation. Moreover, in this cluster, many parties exceed the national three per cent threshold and there is strong electoral competition.

Mid-size cities and industrial agglomerations are concentrated in cluster SD2, showing the highest level of employment in Industry. Lifestyle stability and homogeneity, civic duty and high-income levels determine the highest turnout levels in both electoral rounds.

Industrialised Central-Northern areas belong to super-district SD3. It represents an ‘old’ part of Italy; among super-districts, SD3 exhibits the highest aging index and percentage of residential buildings constructed before 1945. Organisational interactions existing in industrialised areas and their need for changes in modern times might be among the factors influencing electoral participation. Indeed, this super-district is characterised by the smallest turnout variation and the highest electoral closeness in both rounds.

The super-district SD4 represents the most developed services and economic hubs of Italy. It includes the largest cities, i.e. Rome and Milan. Its residents show higher income levels and they are more educated with respect to residents in other super-districts. The electoral competiveness was quite high in this area. Indeed, the difference between the first two candidates was 5.3 per cent and 8.8 per cent in 2018 and 2022, respectively. This competiveness level contributed to a turnout level aligned with national outcomes.

Cluster SD5 encloses electoral districts around Naples. In SD5, population density reaches its maximum value, while aging index reaches its minimum value. Given that young people vote less, a low turnout might be expected. Additionally, all economic indicators highlight severe employment and economic issues. For example, the employment rate is 30 percentage points below the maximum level over super-districts and the female employment rate is half of the corresponding male employment rate. Moreover, the average income is below the national level. The perception of such difficulties and their continuous worsening lead voters to have a deep distrust of political parties. In both electoral rounds, the difference in votes between the first two parties exceeds 25 per cent; SD5 represents districts without any electoral competition. Simultaneously, voters did not positively evaluate the incumbency. Despite an increase in the number of parties exceeding the three per cent threshold, no valid political alternative was proposed. Hence, other voting mechanisms were not activated, see Stiers (2022). A consequence of these factors is that the Naples super-district shows the largest voter turnout loss.

Clusters SD6 and SD7 are both located in the South of Italy and share many socio-economic and electoral features. The super-district including the Italian main islands, Sicily and Sardinia, shows the worst economic performance among the seven super-districts. Low levels of both education and income strongly characterise these two groups of electoral districts. In both super-districts, the percentages of male candidates born in those areas are close to maximum values. In 2022 elections, the continental super-district shows the lowest percentage of parties exceeding the three percent threshold, without any significant variation; it indicates a difficult political innovation or change. These super-districts delineate areas where trust in old-fashion political parties reaches its minimum and consequently causes decreasing and less significant electoral participation. Clusters SD6 and SD7 are those mostly contributing to the large national turnout variation.

To summarise, unequal electoral participation theories introduced in Lijphart (1997) seem confirmed. Indeed, higher income-more educated residents in Central and Northern super-districts show a greater participation rate. The relationship between turnout and electoral closeness clearly emerges, too. Low voter turnout, its decreasing trend and electoral competiveness are simultaneously present in five super-districts. This claims for further evaluations of turnout trends, electoral closeness benchmarks and endogeneity, etc.

With regard to party environment, super-districts where parties attempted to address nowadays issues related to, for example, gender equality and territoriality display a lower turnout decrease. Clusters SD1-4 increased either the percentage of female candidates or the percentage of candidates born in their areas and succeeded in restraining the turnout decrease. Compared with SD6, parties running in SD7 have paid slightly more attention to the selection of candidates. The number of parties exceeding the three per cent threshold is generally increasing. This is the consequence of both mistrust in old political parties and missing newer valid political proposals. Statistical analyses on the relationship between candidates’ individual characteristics, parties and turnout are out of the scope of this work.

As already mentioned, in our application, AP standard setting, Frey and Dueck (2007), generated seven almost spatially equal clusters suitable for a cartographic representation. Anyway, the R package apcluster implements several options enabling a more advanced control over the number of resulting clusters. Consequently, users may determine themselves the suitable degree of aggregation.

4. Conclusions and further work

The objective of this study is to propose a classification of the Italian electoral districts by means of affinity propagation. Being derived from travel-to-work flows and socio-economic indicators, the classification might be used to illustrate contextual analysis comparisons. Referring to Italian 2018–2022 parliamentary elections, the Main Map depicts turnout variations in super-districts, i.e. socio-economic homogeneous aggregations of electoral districts. Super-districts prove to be useful spatial units for assessing the existence of relationships between trends of turnout and voting determinants.

Depending on aggregated data, the proposed methodology can be applied in many different contexts. Moreover, the two-stage affinity propagation approach is flexible enough to be adapted to other case-studies: sets of socio-economic indicators, types of flows linking electoral districts and number of clusters are few parameters that could be manipulated in order to derive meaningful spatial units. Further research should concern the contiguity of super-districts and the opportunity to derive hierarchical classifications.

Software

R 4.3.1 software was used for data manipulation, (R Core Team, 2023). R library apcluster Bodenhofer et al. (2011, 2023) was used for aggregating districts. QGIS version 3.22.13-Białowieża (QGIS Development Team, 2022) was used for cartogram creation.

Data

Italian electoral districts boundaries and socio-economic indicators are available at https://www.istat.it/it/archivio/basi+geografiche+collegi+elettorali.

Italian travel-to-work flows are available at https://www.istat.it/it/archivio/139381.

Italian electoral data are available at https://elezioni.interno.gov.it/.

Income data is available upon reasonable request.

Acknowledgements

Partial financial support was received from Eurostat Grant Agreement N. 882021 2019-IT-Subnational and PON GOVERNANCE 2014-2020 Informazione statistica territoriale e settoriale per le politiche di coesione 2014-2020.

Istat is not responsible for any views expressed in this paper.

The authors thank the referees for their helpful comments that improved the quality of this paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by Eurostat [Grant Number N. 882021 2019-IT].

Notes

1 In 2022, the only additional factor negatively influencing political participation could be represented by the COVID-SARS-2 pandemic period.

2 Nomenclature of Territorial Units for Statistics.

3 Sicily is an exception due to already low electoral participation.

4 https://www.normattiva.it/uri-res/N2Ls?urn:nir:stato:legge:2017;165 (in Italian).

5 The latest available. The Origin–Destination matrix stems from population Census.

6 Without a-priori defining the number of clusters.