In this section, we present our estimates of the impact of large-scale, in-person elections on the COVID-19 pandemic spread in four steps. First, we examine the impact of elections on the growth in new infections. Second, we estimate the electoral effect on the growth in hospitalizations. In the third step, we inspect heterogeneity in the pandemic spread. In the fourth and final step, we shed light on the mechanism of viral spread by estimating the impact of elections on physical mobility in the election week.
New COVID-19 cases
First, we consider the effect of elections on new infections. Figure 2 reveals a rapidly accelerated growth in new COVID-19 cases in voting relatively to non-voting constituencies after the second round of the 2020 Senate elections. Panel A plots the average 14-day growth rates in new COVID-19 cases in absolute values across voting and non-voting constituencies using municipality-level data with 6,259 units observed within − 28 to + 56 days around elections (N = 532,015). Panel B plots coefficients βj obtained from Eq. (1) which represent the estimated differences in the 14-day growth rates across voting and non-voting constituencies in every day in the inspected window around elections. The coefficients are multiplied by 100 to show percentage difference in the growth rates.
The coefficients in Panel B indicate that the 14-day growth rate in new cases started accelerating approximately 1 week after elections. This period corresponds exactly to the median incubation period for COVID-19 augmented by a lag of 2–3 days, which are likely associated with voters seeking testing and acquiring test results. The difference in the growth rates continues rising in the second week after elections and becomes significant at the 5% level 14 days after elections. The effect is most pronounced and markedly significant in the third week after elections.
Figure 8 in the Appendix shows a very similar pattern when the inspected outcome in Eq. (1) is the 7-day growth in new COVID-19 cases. The figure reveals that the 7-day growth starts accelerating 1 week after elections and is significantly faster in voting constituencies in almost the entire second week. The acceleration is slower in the third week, as the 7-day growth rate already takes the elevated growth rates on the break of the first and second weeks as the reference values.
In Table 2, we estimate the average acceleration in the 14-day growth rate in new COVID-19 cases in the third week after elections using Eq. (3). In the most parsimonious specification in column (1), we find that the 14-day growth in new cases is 24 percentage points higher in voting compared to non-voting municipalities. Relatively to the average 107% growth in all municipalities, new cases thus grow 23% faster in voting municipalities. The estimates are barely affected in column (2) where we account for (observed and unobserved) time-invariant municipality-specific factors by adding municipality fixed effects. They also remain very similar in column (3) where we control for municipal-specific time-varying pandemic situation 14 days earlier. The estimates for the interaction term in all columns are significant at least at the 5% level.Footnote 11
In absolute terms, the excess number of new infections generated by elections can be calculated by multiplying (i) the estimated acceleration in the growth rate of new cases in voting relatively to non-voting constituencies in the third week after elections (which reflects the previous 14 days since the first signs of growth rate acceleration), (ii) the average prevalence of 1,231.56 cases per 100,000 people observed 1 week after elections (before the appearance of the election effect), and (iii) the population in voting constituencies. The product of these numbers corresponds to 14,858 additional cases when we binarily classify the three largest cities as treated. A more conservative estimate, which approximates the population in voting constituencies in the three largest cities by the share of voters living in their voting constituencies, would suggest excess 10,791 cases. Finally, the most conservative estimate which entirely omits the contribution of the largest cities suggests the excess of 8,692 new cases.Footnote 12
As a robustness check, we estimate the number of excess cases also from Table 10 in the Appendix, in which we estimate the differences in 28-day growth rate in new infections across voting and non-voting constituencies 28 days after elections using simple cross-sectional OLS. If the estimates are multiplied by the population in voting constituencies and pandemic prevalence on the day of elections, we estimate excess 15,877 cases due to elections when the three largest cities are classified as treated. The similarity of the estimates with respect to the previous figures suggests that most of the additional cases arose exactly during the second and third weeks after elections.
We highlight two additional observations in Fig. 2. First, we note that the election effect fades away in the fourth and later weeks after elections. This is in line with our intuition that social interactions on the election day produce a one-time boost in the prevalence of new cases, but after it is reflected in statistics, the pandemic continues to grow at equal rate in voting and non-voting constituencies, although from an elevated base in voting constituencies. Second, we point out that there are no significant differences in the growth rate in new infections across voting and non-voting constituencies at any date prior to elections. This lack of pre-trends strongly adds credibility to the causal interpretation of our findings.
We continue by estimating the impact of elections on the growth in hospitalizations. The advantage of this outcome is that it is far less dependent on country-specific standards in detecting and reporting COVID-19 cases. It can therefore help us validate if the acceleration in new cases is merely due to increased testing in voting constituencies.
Figure 3 shows that active hospitalizations grow significantly faster in the third week after the second round of Senate elections in communes with higher shares of population from voting constituencies compared to communes with fewer eligible voters. The figure namely visualizes coefficients δj from Eq. (2) estimated using commune-level hospitalization data with 206 units within the − 28 to + 56 days frame around elections (N = 17,510). The coefficients are multiplied by 100 to show percentage differences in hospitalization growth rates.
A visual inspection of Fig. 3 suggests that the growth in active hospitalizations started accelerating 12–14 days after the second round of Senate elections and became significantly higher at the 5% level in communes with a higher share of eligible voters in the third week after elections. This pattern, if anything, suggests only a short delay in the growth acceleration in hospital admissions relatively to the acceleration in new infections. Figure 10 in the Appendix yet reveals that in the corresponding time period at the end of October 2020, around 50–60% of COVID-19 cases who were admitted to hospital in the Czech Republic were first tested positively for COVID-19 only after hospitalization. The figure thereby partially explains why the dates of appearance of the accelerations in new detected infections and hospitalizations are not very distant from each other.
In Table 3, we quantify the average acceleration in the hospitalization growth in the third week after elections using Eq. (4). In column (1), we report coefficients from the most parsimonious specification without commune fixed effects indicating that hospitalizations grew 62 percentage points faster in communes with 100% of population residing in voting constituencies compared to communes with zero population eligible to vote. The coefficient is significant at the 5% level. Relatively to the average 170% growth in new hospitalizations in all communes, new hospitalizations thus grow 36% faster in fully voting compared to non-voting communes. The coefficients are slightly higher in magnitude in columns (2) and (3), when we include in the model the commune fixed effects and the time-varying controls for earlier pandemic situation. The estimate in column (2) in not significant at the conventional levels (p = 0.106) and the estimate in column (3) is significant at the 10% level (p = 0.074). We argue that the standard errors are relatively high due to the relatively lower granularity of hospitalization data compared to the data on new infections observed at the municipality level.
If we take the results presented so far together, we note that the acceleration in hospitalization (around 36.4–46.4% relatively to the average growth rate) is somewhat higher than the acceleration in new cases (around 22.5–26.3%). This could be expected when citizens are reluctant to get tested or when testing facilities are overloaded and citizens are not tested until the disease progresses into a more severe phase requiring hospitalization. In both scenarios, the temporarily interlinked nature of the two accelerations rules out that estimated effect of elections on viral spread would be merely due to increased testing in voting constituencies. In Fig. 11 in the Appendix, we provide additional evidence against this hypothesis by inspecting the differences in average 7-day positivity rates across voting and non-voting constituencies using event study specification from Eq. (2) and commune-level data. The figure indicates the positivity becomes around 1 percentage point lower in constituencies with 100% of voting population in the second week after elections compared to constituencies with zero eligible voters. The effect is not statistically significant at the conventional levels. It is also not large in magnitude especially when expressed relatively to the average positivity of 32.8% observed in the second week after elections. In sum, our evidence suggests that changes in the intensity of testing likely play a limited role in the observed pandemic acceleration.
Next, we examine heterogeneity in the pandemic acceleration with respect to the observable characteristics of the infected cases and municipal population.
We start by examining pandemic acceleration in new COVID-19 infections separately in population younger and older than 65. In panels A and B in Fig. 4, we plot coefficients βj obtained from Eq. (1) when the 14-day growth rate in new infections is calculated only using cases younger and older than 65, respectively. In Panel A for cases younger than 65, we depict the differences in the pandemic growth rates across voting and non-voting municipalities that are very similar to what we observe in Fig. 2 for the whole population. Approximately 1 week after the elections, the 14-day growth rate in new cases starts accelerating in voting relatively to non-voting municipalities. The acceleration is discernible during the second and third weeks after elections, and later the election effect fades away. Statistically, the coefficients are significant at the 10% level around the peak of the acceleration. In Panel B, we observe a very different pattern for the population older than 65. There is essentially no discernible acceleration in the pandemic growth in voting relatively to non-voting constituencies within the entire examined period of 2 months after the elections.
We interpret the differential acceleration as evidence consistent with strategic risk-avoidance by older voters (Dave et al. 2020b), for whom COVID-19 represents a major risk of hospitalization and dying (Williamson et al. 2020). In theory, the reasons for the differential impact might be that either older cohorts are more cautious in taking preventive measures and following social distancing protocols or they are simply more likely to abstain from elections. Since elections are anonymous and exit polls were not conducted, we do not observe turnout by age groups. In Fig. 12 in the Appendix we however plot the associations between municipal share of population older than 65 and total turnout in 2016 and 2020 regional elections, which were held together with the first rounds of the Senate elections.Footnote 13 We find that in the 2016 elections, a 1% higher share of population above 65 was associated with 0.511% higher turnout. In the 2020 regional elections, held 1 week before our natural experiment, the estimated association was four times lower in magnitude and insignificant at the conventional levels if one accounts for municipal population size. The weaker association in 2020 together with the pandemic acceleration absent in population above 65 point towards increased absenteeism in elections by older cohorts.
In Fig. 5, we continue examining heterogeneity in the pandemic acceleration due to elections with respect to socio-economic conditions in municipalities. In particular, we proceed by dividing the sample of all municipalities into halves according to the median values of municipal employment and the median share of individuals with at least secondary education, respectively. Then, we estimate Eq. (1) using each of the reduced samples and the 14-day growth in new infections as the outcome variable.
In all panels in Fig. 5, we observe a discernible acceleration in the pandemic growth rate in voting compared to non-voting constituencies peaking in around the third week after elections. We note that the acceleration is significant at the 5% level only in municipalities with below-median employment and below-median share of individuals with at least secondary education. When we test the equality of the acceleration across municipalities with below- and above-median levels of employment, we however do not find statistically significant differences. On the other hand, we find the acceleration significantly higher in municipalities with below-median share of individuals with at least secondary education compared to municipalities above the median. We interpret our results as consistent with the literature suggesting that socio-economic factors play an important role for the speed of the pandemic spread and its mitigation. In this literature, for example, Wright et al. (2020) show that regions with higher economic endowment are more likely to comply with anti-pandemic measures.Footnote 14
Physical mobility and social interactions
Finally, we examine the mechanism of viral spread by asking if elections are associated with spikes in physical mobility, and if so, what specific mode of social interactions might have contributed to faster pandemic spread on the electoral days.
In Fig. 14 in the Appendix, we start the analysis by visualizing the association between the first round of Senate elections and country-level mobility indices from Apple. Panels A, C and D namely plot mobility indices for Thursdays, Wednesdays and Tuesdays (non-electoral days), respectively, in a range of − 10/+ 5 weeks around the elections. Panel B plots the mobility indices for Saturdays (electoral day) within the same time frame. In all panels, the figure shows a generally declining trend in all examined types of mobility (walking, driving, transit), consistently with the expectation that people were continuously limiting mobility in the face of the progressing pandemic and government restrictions. At the same time, Panel B shows a pronounced spike in mobility on the electoral Saturday. The magnitude of the spike should be however interpreted as suggestive, as any day-specific shocks, such as favourable weather conditions, might have elevated mobility on the electoral day. At the same time, the data rely on users of Apple devices, which might be more strongly represented in larger cities with higher turnout.
In Table 4, we provide evidence from a representative survey “Life during the pandemic” which has been following a panel of Czech households since mid-March 2020. Respondents from across different districts were bi-weekly asked about the frequency of various social activities which they had participated in each of the previous two weeks (such as in-person shopping, family visits, visits to restaurants and pubs, group holidays, attendance in large public events). The table shows estimates from random effects multinomial ordered logistic regressions, in which the outcomes are categorical variables representing the frequency of particular activities. In fashion of Eq. (4), the independent variable of interest is the interaction between the dummy for the electoral week and the share of the district population living in voting constituencies.
The table indicates that the probability that respondents travelled at least once by crowded public transport increased 6.1 percentage points (10.9%) in districts with 100% of population from voting constituencies in the week of the second electoral round compared to districts with no such population. The estimate is significant at the 10% level. It is also quantitatively feasible given the turnout of around 16.74% in the second electoral round. At the same time, we estimate that none of the other surveyed activities were statistically more likely to be carried out by respondents in the election week in districts with higher shares of population from voting constituencies.
The null results for additional social activities are supported by Table 5, in which we use Google mobility reports to examine the effect of elections on physical mobility at six different general types of locations (retail and recreation facilities, groceries and pharmacies, parks, transit stations, workplaces, and residential areas). Using a variant of Eq. (4), we find that elections are not significantly associated with higher mobility at any of these locations. On the other hand, they are significantly linked with a shorter length of stays of the tracked devices at residential locations on the electoral day (column 6). The effect is significant at the 5% level.
In sum, our estimates suggest that in-person elections are linked with elevated mobility, which does not seem to be related with higher social interactions in any of the examined locations outside polling stations. One should yet remain cautious in interpretation, as it is possible that the examined survey omits an important category of activities or that Google does not track a category of locations that were key to viral spread. In addition, it is important to remember that many of primary infections from the election days produce secondary and tertiary cases in voters’ households and workplaces, contributing to the total acceleration in new cases due to in-person voting.