Estimating Crowd: Electoral Adjustment and Spatial Effects of Turnout During Covid-19 Pandemic

3.1 Administrative Adjustment: Polling Site’s Population Size

Under the absence of convenient (or distance voting), engineering polling stations by enlarging their number, so lowering the size of voters in each polling site than usual elections, is a plausible electoral adjustment. Though it looks simple, reducing population size in a polling station arguably pre-empts the overcrowding of voters on election day as a crowd is a critical issue during the infectious pandemic. Though most studies agree with the negative effect of population size on turnout (Geys 2006, 642), the competing debate remains in the homogeneity thesis. They argue that a smaller size of a population that generates a homogeneous society increases turnout as group solidarity and feasible linkage between politicians (or political leaders) and voters foster people to vote (Kostadinova and Power 2007). However, in a time of crisis, population size relates to the extent to which voters expect their health risk, and people are cautious when an overcrowded polling site is expected. All these institutional arrangements and adjustments return to the theoretical premise of the cost of voting. Drawing on this notion, this model hypothesizes that:

3.2 The Proximity of Crisis: Covid-19 Infections

Following Downsian rational-choice model, the proximity of crisis (i.e. the extent to which Covid-19 infections occur in the neighboring area) arguably affects political participation. This mechanism is similar to the population size model. As pioneered by Downs (1957), the rational-choice model has been employed extensively to explain the behavioral premises of voter turnout. This model, nevertheless, is hardly ever linked with institutional factors such as administrative arrangements of how an election is held. It primarily suggests the idea of the cost of voting, or “calculus of voting” derived from the utility function of rational-choice theory (see Aldrich and Jenke 2017). In addition, the role of information distribution is also under this model (Feddersen and Pesendorfer 1999). For instance, though Morton et al. (2015) estimate the effect of exit poll information on the decreasing turnout, they suggest that exit poll information also increases bandwagon voting. Similarly, in the case of the Covid-19 pandemic, a voter is expected to stay at home when any type of distance voting is absent, as people hesitate when surroundings are infectious. Based on the cost of voting-related health risks, we may propose a pandemic model of turnout as follows.

3.3 Close Election: Political Mobilization

An election is somehow a political contest among candidates (or parties) that situate them to compete by mobilizing voters to vote for their electoral gain. Other than pandemic-related arguments, the political mobilization model may give us an enduring theory of voter turnout, either in normal or crisis circumstances. Political mobilization theory relates to the idea of a close election. By close election, it means that when there are two competing candidates who contest competitively, indicated mainly by a small vote margin, the likelihood the voters cast their votes is higher than in an election under uncontested candidates or non-competitive race (Cann and Cole 2011; Grofman 1995; Indridason 2008; Simonovits 2012). Close election situates candidates to mobilize voters and media to report more coverage regarding the contesting candidates, which boosts people to vote. Based on their massive field experiment, Green and Gerber (2019) documented several campaign approaches of get-out-the-vote techniques that foster turnout, including mail, postcard, and in-person visit (canvassing) to increase turnout (Gerber and Green 2000, 2005).^[2] As a close (or competitive) election indicates the extent to which the candidates mobilize voters, this article hypothesizes that:

Lastly, demographic characteristics have been crucial variables that possibly confound those models. Demographic characteristics, such as gender, age, and education, have always been inevitable variables in explaining voter turnout (Plutzer 2017). For instance, a high level of education and literacy rate is linked with electoral participation Diwakar (2008). Additionally, age potentially plays critically in a time of pandemic where elder voters are susceptible to such a circumstance. In short, we also need to control for these variables should we examine those three models.

5.1 Data

This study examines data from the Surabaya mayoral election in Indonesia, held on December 9, 2020. Turnout in previous elections – 51 % in 2015, 44 % in 2010, and 48 % in 2005 – was consistently lower than in the 2020 pandemic election, which reached 53 %. This increase is counterintuitive given the global health context: the World Health Organization (WHO) declared a Covid-19 outbreak on January 30, 2020, and officially categorized it as a global pandemic on March 11, 2020. Covid-19 vaccines were not available until late 2020 and were distributed globally only in 2021. Therefore, elections conducted between March 2020 and early 2021 occurred during the peak of the global crisis – a critical period of the Covid-19 pandemic.

Analyzing turnout in elections held during this critical period allows the study to contextualize the 2020 Surabaya mayoral election. Examining turnout trends across countries provides empirical insights into global patterns of electoral participation during the pandemic and positions Surabaya’s case as distinctive. To that end, this study utilizes turnout data from 2,390 national elections compiled by International IDEA. Among these, 24 countries held elections between March 2020 and January 2021. With the exceptions of Romania and Anguilla, most of these elections were held on or after January 1, 2021 – having been postponed from earlier dates that would have coincided with the pandemic’s most critical period. It is important to note that International IDEA records only national-level elections; local elections, such as the one held in Surabaya, are not included.

Drawing from these “pandemic elections,” the study compares turnout in two electoral cycles for each of the 24 countries: one held during the critical period of the Covid-19 pandemic and one just prior to it. As shown in the boxplot in Figure 1, the global average turnout during the pandemic period was lower than in the immediately preceding elections. This makes the increased turnout in Surabaya particularly noteworthy. The accompanying dot plot illustrates that the number of countries experiencing a decline in turnout (i.e. percentage differences below the zero line), such as Canada, Israel, Peru, Argentina, and Uganda, significantly exceeds the number of countries – such as Mexico, Morocco, Iceland, and Nicaragua – where turnout during the pandemic exceeded pre-pandemic levels (i.e. percentage differences above the zero line).

Figure 1:

Differences of turnout across elections held before and during the critical period of Covid-19 pandemic.

Therefore, investigating an increase of turnout in 2020 Surabaya municipal election provides an exemplary notion on how we explain an increase of turnout during a time of crisis. In this regard, the election was contested by two candidates who, to some extent, made the case suitable to examine the models. Furthermore, credited as the second largest city after Indonesian capital Jakarta and as the capital of the second largest province East Jawa in Indonesia, Surabaya is a melting-pot megapolitan inhabited by more than three million people. With Indonesian threshold of poverty rate ($2 USD per day income), only about 4–5% Surabaya households that are at poverty level in 2020 based on Indonesian Statistics Bureau. This figure is considerably low relative to other cities or global poverty (about 9 % or more). In addition to unavailable village-level economic data, this low economic disparity helps anticipate empirical examinations from economic confounders.

Other than data availability and research feasibility, the municipal administration provides Covid-19 aggregate data on a daily basis regarding its 154 village-level units. This data juxtaposes with mayoral-election data provided by Indonesian General Election Commission which is also at village-level units. The third data set on demographic characteristics complement the two key datasets. As this research also fielded data collection on the ground by visiting mayoral and sub-district offices for village-level data and validation, these are the most updated data in 2020 and 2021.

In addition, in the case of urban villages as aggregate-level units, an urban village (kelurahan) is a small unit of electoral constituencies that may uncover such spatial patterning of social-political interactions. It is because neighborhood politics and social interactions occur among voters at the village level where polling sites are administered, Covid-19 infection cases are registered and reported, and the votes are contested.

Figure 2 shows the maps of the spatial distribution of Covid-19 infection cases on the right-hand side and turnout on the left-hand side. We see that the voter turnout is relatively high, indicated by 50 % or more turnout, in the northwestern villages, while the distribution of Covid-19 infection cases is also relatively low, marked by 10 or more minor cases, in the region. At a glance, these maps give us a sense that the relationship between turnout and the Covid-19 pandemic is negatively associated spatially. As this interpretation might be premature, we need to investigate further the spatial models examining the distribution of Covid-19 and turnout elaborated later.

Figure 2:

Voting during the Covid-19 pandemic in 2020 Surabaya mayoral election.

5.2 Variables

Voter turnout, the dependent variable, is defined as the percentage of total votes cast relative to the number of registered voters. While some scholars use the voting-age population as the denominator, the available village-level data include only official records of registered voters. Moreover, the number of registered voters is presumed to closely approximate the voting-age population at the village level.^[3]

To examine the cross-sectional models, several explanatory variables are developed. The first explanatory variable – the average number of voters per polling station – captures the relationship between the total number of polling sites and the total number of registered voters. This variable is central for approximating the electoral adjustment made in response to the Covid-19 pandemic. In the Indonesian context, the maximum number of voters per polling station was reduced from 800 in pre-pandemic elections to 500 stations during the pandemic, reflecting an institutional effort to reduce crowding on election day.

For the political mobilization thesis, approximated by close election (competitiveness), the other independent variable addresses the difference between the number of voters gained by the two candidates. Thus, the smaller the contrast of the votes indicates a high level of competitiveness, approximating extensive mobilization in the given village. Relatedly, the proximity of Covid-19 refers to the rise of the number of infection cases one week before the election. The increase in infection cases sets the alarm to the voters not to take their health risks by attending polling sites.

The analysis includes several control variables that may confound the relationship between explanatory factors and voter turnout. The age composition of each administrative unit is considered, as villages with a higher proportion of older residents – who are more vulnerable to Covid-19 – may exhibit lower turnout rates compared to those with a younger demographic profile. Specifically, the proportion of inhabitants aged 50 and above is calculated for each village.

Gender composition is also included as a control variable, measured by the ratio of males to females. Villages with a greater proportion of male residents are theoretically expected to exhibit higher turnout. Additionally, educational attainment is accounted for, drawing on the resource model of political participation, which posits that individuals with higher levels of education are more likely to vote. The variable is operationalized as the proportion of residents in each village who have attended higher education.

Finally, the number of religious places of worship is included, with mosques being the predominant type across villages. These institutions often function as centers of political socialization and community engagement, which can facilitate political mobilization and increase voter turnout. Descriptive statistics for all control variables are presented in Table A1 of Appendix A.^[4]

5.3 Methods

My empirical strategy heavily employs spatial econometrics by addressing data on hundreds of urban village-level units in Surabaya. The underlying assumption of spatial econometrics is that the units are not independent as social interactions are inevitable (see Anselin 2007; Anselin and Florax 1995). Thus, spatial approximation, known as spatial weight matrix, is necessary to measure the extent to which the interactions occur and affect the resulting behavior of voter turnout. In this regard, I employ first-order queen contiguity for the spatial weight matrix (W). The type of data set of 154 village-level units that share their borders for relatively short distances situates this paper to apply first-order contiguity rather than second-order contiguity. Moreover, queen contiguity here is aimed to have all the surrounding village neighbors, even those that share a small magnitude of borderlines.^[5]

Based on that basic logic, assessing the randomness of the observations scattered across spaces and locations is imperative. In many instances, the spatial patterning is a result of demographic groups living in the nearby area (see Cho and Gimpel 2010). Autocorrelation, computed mostly through Moran’s I statistics, is a stochastic method that gauges the spatial randomness of observations, i.e. whether the units’ variable of interest is spatially clustered/structured or random. In other words, once we find a significant positive coefficient of the Moran’s I statistics, we may temporarily conclude that voter turnout is spatially correlated (clustered), i.e. a high (or low) turnout in one village is associated with a high (or low) turnout in its surrounding villages, and otherwise. Global Moran’s I Statistics is computed as follows.

where i and j index the spatial units of which there are n, w _ij is an element of a row standardized spatial weights matrix, x is the percent of the turnout, and μ is the average turnout percentage in the sample.

While we are equipped with the statistics to assess spatial patterning across all the village-level units, the Local Moran’s I or local indicators of spatial autocorrelation (LISA) is able to show the spatial clustering of turnout between the neighboring units. This test statistic is computed as follows.

where z is the mean-deviated form of the percent of the turnout in particular villages.

Once we find a positive and significant result in Moran’s I statistics, i.e. an indication of spatial patterning of voter turnout, we need to specify the spatial autoregressive models. The models estimate the spatial effects of the variable of interest or the error relating to any omitted (unobserved) variables that are spatially correlated with voter turnout. As I theoretically suggest that turnout effects are local rather than global, i.e. turnout in an urban village affects more its neighboring villages rather than all villages in Surabaya, Spatial Durbni Error Model (SDEM) is pursued through the following equations.^[6]

and

where y is a vector that represents the dependent variable of voter turnout, W is an n × n spatial weights matrix, X is the explanatory variables with β as the coefficients, and ε is the error term. Meanwhile, we seek whether there is an error that relates to spatial attributes of any omitted variables through the following spatial error equation. λ is the spatial lag parameter to be estimated. Note that a likelihood ratio test has been conducted to investigate whether SDEM should be restricted to a simpler model, e.g., Spatial Error Model (SEM). Appendix provides the test result of the model selection.

The spatial autocorrelation and autoregressive statistics have been applied across disciplines that take spatial effects into account (see, for instance, Baller and Richardson 2002; Voss et al. 2006). In addition, I at first draw Ordinary Least Squares and counterfactual models based on the observed variables to show how spatial models differ from the “regular models” when we specify spatial weighted matrix into our model specifications. Note that this article also reports results from geographical weighted regression (GWR) beneficial to statistically explore the tests. Though geolocational variables specifying longitudes and latitudes (coordinates), either voter’s or polling site’s locations, that are necessary for GWR are not available, centroids (the middle points of an area, i.e. each urban village based on its border line) are able to be computed.^[7] Given the nature of the spatial analyses and the geospatial data, I employ a series of data processing in R, GeoDa, and QGIS software.

6.1 Contagious Turnout

Figure 3 shows that turnout in the election during the pandemic (2020) and the one before the pandemic (2015) are significantly spatially correlated (p < 0.05). Note that both autocorrelations are approximated by the percentage of voter turnout where some squared dots in both plots indicate neighboring correlations of turnout between the neighboring villages. This means that the plots clearly show that a high turnout in one village has a link with a high turnout in the neighboring villages and vice versa. The quadratic plots of autocorrelations in Figure 2 show four relationships between turnouts of neighboring villages, i.e. high-high turnouts on the upper-right, low-low turnouts on the lower-left, high-low turnouts on the upper-left, and low-high turnout on the lower-right areas. In other words, turnout is not spatially random as the Moran I’s statistics indicate the presence of spatial patterning indicated by their positive coefficients (direction of the autocorrelation).

Figure 3:

Global Moran’s I statistics of voter turnout in the pandemic election (December 2020) and the pra-pandemic election (2015).

However, a closer look shows that the autocorrelation coefficient of turnout during a pandemic (r = 0.390) is lower than that in the 2015 mayoral election indicated by r = 0.50. These Moran’s I statistics suggest that turnout has been already contagious, and even at a higher level, before the pandemic. People in regular elections before a pandemic have no constraints to have encountered, and political mobilizations on the ground are arguably more massive than those during pandemic elections. In order to clarify this possible explanation, Local Moran’s I, also known as local indicators of spatial autocorrelation, abbreviated as LISA, gives us a more detailed picture of how the spatial patterning of turnout occurred between the 2020 and 2015 elections.

First of all, the maps in Figure 4 show a similar spatial clustering of voter turnout between the 2020 and 2015 mayoral elections, indicated by think-colored neighboring villages having spatially significant autocorrelation at a significant level at α = 0.05 and those with a significant level at α = 0.1 in thin-colored neighboring villages. With significantly high-high (dark red) and low-low (soft red) autocorrelations, the high voter turnout of 12 villages and the low voter turnout of 33 villages are clustered in the west and in the middle north-east part of the city, respectively, in 2020 election. This cluster of villages in 2020 also occurred 2015 election with an addition in the southwest part of the city. In short, these comparisons of LISA between the two periods of elections suggest that the contagious turnout during a pandemic is not new and the spatial patterning of turnout might have some geographical strongholds.

Figure 4:

Local Moran’s I statistics of turnouts in 2015 and 2020 elections.

Furthermore, suppose we return to the left-hand map of Figure 2. In that case, the clustering of the high turnout in the western part of the city seems to fit the epidemiologist’s expectation of the low infection in the western region relative to the low turnout in the southern part of the city. However, the low turnout in the middle north-eastern part does not meet our low infection expectations. Villages with a low infection rate are expected to have a cluster of high turnouts. This spatial patterning of turnout leads us to investigate further possible factors, including proximity of Covid-19 infections, that situate people to come to polling stations.

Moreover, such electoral participation is even more “infectious” than the virus shown on the right plot. This comparison is not intended to test the theory I proposed earlier. Rather, a comparison between turnout and Covid-19 spatial autocorrelations is to demonstrate the extent to which turnout is contagious. The spatial patterning, or spatial clustering, becomes more evident when we see the spatial mapping of the local Moran’s statistics.

In contrast, as shown in Figures A1 and A2 of Appendix B, the Covid-19 infections before the mayoral election are not spatially clustered clearly with a very low autocorrelation coefficient r (0.004). In other words, conditions in one village are not strongly associated with infections in other villages. Though this is a surprising finding, some suggest that family-based clusters of infections massively took place throughout Indonesia in 2020, including Surabaya city (see Soedarsono 2020). Evidently, as depicted in Figure 6 of the same Appendix, we do not find spatial patterning of the Covid-19 infection across villages as massive as those in turnout. Thus, it is more critical to investigate how ‘traditional’ and new pandemic-related variables, juxtaposed with the spatial effects, affect turnout during the pandemic rather than merely to seek how Covid-19 virus affect turnout.

Figure 5:

Spatial durbin error models for turnouts between 2020 and 2015 Surabaya municipal elections.

Yet, in what mechanisms the turnout in both elections is spatially structured, and what are the drivers that situate voters to eagerly turn out to vote in a time of such a crisis? The following results elaborate on these questions.

6.2 Spatial Models of Turnout

As mentioned earlier, I theoretically propose that spatial effects of turnouts are local. This means that turnouts in one spatial unit (i.e. an urban village in Surabaya) affects turnout in its neighboring villages, rather than globally affecting all urban villages across in the city. Thus, modeling the effect through Spatial Durbin Error Model (SDEM) is theoretically reasonable. Additionally, the likelihood test between SDEM and Spatial Error Model (SEM), reported in Appendix, also provides statistical reason to maintain SDEM though I did cross-sectional examinations through multiple models, as reported in Appendix, before jumping to examine spatial dimensions in the model specifications.^[8]

Following on the theoretical proposition and the statistical test, Figure 5 displays how SDEM-based spatial effects of turnouts occur in 2015 before-pandemic and 2020 during-pandemic elections. In this regard, I apply variables generated in the 2020 election for 2015 turnout as counterfactual models, i.e. how such variables, especially the Covid-19 pandemic and electoral adjustment, occurred in the 2015 election, which in fact did not occur.

Figure 6:

Spatial classification of local coefficients of geographically weighted regression of the effects of Covid-19 cases, competitiveness, and polling population on turnout in 2020 Surabaya election.

Lambda in Figure 5 highlights the extent to which local spatial effects occurred in before and during pandemic election. Its statistical significances at p < 0 .05 suggest that turnouts are spatially correlated in both elections, which is consistent with H4 on spatial effects of turnout. This means that turnouts as a proxy of electoral participation in one village are spatially correlated with turnouts in neighboring villages. A closer look, however, signifies that the spatial magnitude is higher during 2020 pandemic election (0.56) than before pandemic in 2015 election (0.48), suggesting that local neighboring effects of turnout is higher during the pandemic. In other words, turnout is more “contagious” during pandemic than in 2015.

If we break the models down into the specifications of the variables, population size at the polling sites in the election during a pandemic convincingly explains the voter turnout (p < 0.05). We are nevertheless uncertain in 2015 if the pandemic occurred though the estimates show the same direction as in 2020. Its negative coefficients, especially in 2020 models, meet our theoretical expectation that turnout in a large polling site’s population size is low when the population size at the polling site is relatively large. An increase in population size at the polling site contributes to a decrease in turnout, all else equal, since people are cautious when overcrowded is expected as the result of their neighboring interactions. Note that the tests do not find such significant effects of the electoral adjustments in the 2015 models, which makes a lot of sense if we refer to our theoretical expectation of the cost of voting during a pandemic: the importance of assessing physical contacts and overcrowded in a polling site during the pandemic.

Meanwhile, Covid-19 proximity in the counterfactual election in 2015 and pandemic election in 2020 shows negative coefficients. This result, at a glance, follows the theoretical expectation that turnout decreases when the number of infection cases in the neighborhood is high. Nevertheless, these estimates are not statistically indistinguishable. Thus, we are not sure whether voters follow such a theoretical direction. Accordingly, political mobilization, approximated by the level of competitiveness (close election or the margin of votes between candidates), even does not show any sign of our expectation given the contradictory directions of its coefficients across the two elections and its non-significant results at α = 0.05. These results suggest that pandemic proximity and political variable do not explain why voters vote (or not) during the pandemic.

Different from the two focal variables, the proportion of males and the proportion of people who attend higher education in each village are statistically significant across models. These variables signify that they explain turnout in both electoral circumstances. Their positive coefficients confirm the resource model of turnout, where men and people with higher education are more likely to vote. Importantly, people who are 50 years old and older are only statistically significant in the 2020 models, which makes more sense in a time of the pandemic. Their negative coefficients maintain that villages with a higher proportion of elders tend to have lower turnout during the pandemic. Nevertheless, these cross-section models may be necessary to elucidate turnout during a pandemic, but they are not sufficient to explain the rise of turnout during the pandemic. Thus, we need to look at how spatial effects might operate in a turnout along with the observed variables.

Table 1 reports how spatial effects of turnout in 2020 election occur across modeling strategies though this article theoretically proposes and tests the significance of the Spatial Durbin Error Model (SDEM).^[9] Overall, the statistical significances, the point estimates, and standard errors differ (reduced or corrected) across the models are “corrected” from the one to another models when we specify weighted spatial matrix into the model specifications. Male proportion is no longer consistently significant across the models, and most of the estimates of the variables are reduced. Other than the proportion of voters who attend higher education, the result highlights that the electoral adjustment during the pandemic, approximated by the reduction of the polling site’s allocated population size, remains statistically significant. But again, they are corrected to be smaller effect sizes.

As an exploratory analysis, Figure 6 presents maps of the local coefficients for three key explanatory variables – Covid-19 cases, electoral competitiveness, and polling site population – estimated using Geographically Weighted Regression (GWR). The model specification follows a multiple linear regression framework, consistent with the SDM and SDEM models reported in Table 1. The bandwidth for geographic weighting was determined through cross-validation, and the spatial coordinates used were centroids derived from the Surabaya city shapefile.

The results reveal a diverse spatial pattern in the relationship between the three focal variables and voter turnout. Figure 6 depicts the spatial distribution and directionality of these local relationships using a graduated red color scale, where deeper opacity indicates stronger and typically more positive associations. In line with our theoretical expectations, negative relationships are anticipated: lower Covid-19 case counts, smaller voter populations at polling sites, and greater electoral competitiveness (i.e. narrower margins between candidates) are expected to correspond with higher turnout.

A closer inspection of the maps indicates that the spatial variation in the coefficients for Covid-19 cases is relatively limited, whereas the relationship between polling site population and turnout exhibits the highest degree of spatial dispersion. This wide variation underscores the localized effect of polling site population on turnout, a finding consistent with the SDEM model results.

Furthermore, to follow up on the Spatial Durbin Error Model on Table 1, Table 2 shows whether the spatial effects of the variables of interest are statistically significant in terms of their direct and indirect impacts. It shows that we do not have evidence on indirect impacts across the variables since their p-values of SDEM’s impact measures are above 0.1. Yet, the explanatory variable of interest of polling population (0.024 of p-value), along with age and education level, has statistically significant direct impact in terms of the local spatial effects, estimated through SDEM. This means that effects of estimating crowd in the polling site is spatially direct among voters who turned out to vote during the heyday of the pandemic.

Empirically, turnout – driven by how voters estimate the crowds based on polling site population in one village – is a function of turnout in neighboring villages (as computed by the queen contagion weighted matrix) along with their observed variables. In other words, as this result rejects the null of the spatial autoregressive models, turnout is spatially an outcome of both the neighboring observed variable of interest (turnouts) and unobserved variables that might also bear spatial effects since the outcome of turnout is a function of the weighted measure of neighboring turnouts. However, our main takeaway is that the significant result conforms to the theory of contagious turnout. Turnout is a result of neighboring effects where the effects are a spatial outcome of explicit/implicit social interactions of the diffusion theory. In the absence of convenient voting, people who turn out to vote at the polling sites are mainly driven by turnout decisions among their neighbors.

This mechanism conforms to the two-dimensional and multi-directional nature of spatial autocorrelation. Diffusion occurs when the resulting interactions affect each other across multiple neighbors. However, I acknowledge that these models are at the village-level unit and do not reach observations of the individual voters, as many other studies apply spatial econometrics. Yet, the significant results of the spatial autoregressive models detect and indicate individual behavior at the aggregate level since spatial models assume that observations are not independent as the interactions occur between units due to their spatial dimensions.