Voter Elasticity and Political Protest A quantitative study in an American context

(1)

Voter Elasticity and Political Protest

A quantitative study in an American context

Gustaf Westin

Supervisor: Linuz Aggeborn

Department of Government

Political Science C

Bachelor’s Thesis

Autumn of 2020

(2)

Abstract

The purpose of this thesis is to study the relationship between preva-lence of swing voters and the occurrence of political protest. Taking a Rational Choice approach, I hypothesize that fewer swing voters will lead to more protests, because it would incentivize polarizing behavior by political candidates. The hypothesis is tested using protest data from US congressional districts during six months of 2020 as the de-pendent variable, and the concept of voter elasticity as the main inde-pendent variable in a multiple regression analysis, along with various control variables. The results tentatively indicate that the hypothesis is correct, but exhibit high levels of uncertainty, highlighting potential for future research.

(3)

(4)

List of Tables

1 The five most and least elastic districts . . . 23 2 Descriptive statistics . . . 27 3 Regression Results . . . 34 4 Confidence intervals (95%) of regression coefficients for the

elasticity variable in each regression model, 1-8 . . . 36 5 VIF Tests . . . 48 6 Regression Analysis with cluster robust SE s . . . 50 7 Regression analysis in groups, grouped by partisan lean . . . . 51

List of Figures

1 Voter equilibrium according to the Median Voter Theorem . . 11 2 Policy convergence and divergence in a non-polarized &

(6)

1 Introduction

In a television ad titled “So Conservative”1 from his 2018 campaign, Republi-can Governor of Georgia Brian Kemp brandishes his shotgun, uses explosives to blow up and cuts through stacks of papers labelled “government spend-ing and regulations” with a chainsaw, before he is shown sittspend-ing in his large pickup truck which he says he can use to “(. . . ) round up criminal illegals and take ‘em home myself.” He then turns the key in the ignition and says “Yep. I just said that” (Kemp for Governor, 2018).

It is not difficult to imagine what kinds of voters Kemp is trying to appeal to. They probably do not include middle-of-the-road, moderate swing voters or persuadable Democrats. Rather, he is likely trying to enthuse right-wing republican base voters in the hope of increasing turnout among them.

In elections generally, and United States2 _{elections specifically, there is}

often a debate about turnout vs. persuasion – whether the best election strategy is to try to persuade voters (mostly moderate, persuadable centrists) to come over to your side, or try to “fire up the base”, i.e., appeal to your own party’s base in order to increase turnout among low-propensity or non-voters (for a discussion on this debate, see Hall & Thompson, 2020). Kemp, judging by his political ads, seems to have gone with the latter strategy.

This thesis argues that a large part of the reason behind Kemp’s choice of strategy lies in his state’s prevalence of swing voters. In Georgia, Kemp’s home state, relatively few voters swing (vote differently) between elections, regardless of how many do nationally. This means shifts in party support between elections in Georgia tend to be relatively small, and margins of victory relatively narrow.

This dynamic has implications for how to win elections in Georgia. Run-ning a state-wide campaign in which you adopt moderate policies, emphasize bipartisanship and generally pivot to the political center might not be very successful there, simply because the number of persuadable, middle-of-the-road voters in Georgia is very low — most are more or less committed to one party. In contrast, pivoting more to your respective fringe, promoting more ideologically extreme3 viewpoints and using more explosive or even

inflam-1_{Link: https://www.youtube.com/watch?v=5Q1cfjh6VfE} 2_{The US, henceforth.}

3_{The word extreme will consistently be used in a non-normative way throughout this}

(7)

matory language and rhetoric might instead create enthusiasm among voters that may not otherwise have voted at all and thus increase turnout among “your” partisan base voters. Meanwhile, in states or districts that are more elastic, the former strategy may be more successful.

This has implications of its own, in turn. If politicians, prompted by a polarized electorate with few moderate swing voters in their home districts, opt for a strategy of extreme or polarizing political rhetoric and more extreme issue positions, it could arguably have unintended, albeit no less damaging effects on the general political climate. This, in turn, has consequences for a number of different outcomes (Kalmoe, 2014; Piazza, 2020), of which this thesis will focus on one in particular — the occurrence of political protests.

The contemporary US is broadly considered more polarized than it has been in a long time, with exceptionally wide divisions between different sec-tions of society on a large number of political issues - wider than in almost any other similar country (Dimock & Wike, 2020). At the same time, Amer-icans in 2020 have experienced an unprecedented number of public protests and demonstrations. Many of these have been motivated by racial justice, brought on by killings by police such as that of George Floyd and Jacob Blake (ACLED, 2020), but there has also been a number of protests aimed at pandemic restrictions (BBC, 2020).

(8)

1.1 Purpose & Disposition

The purpose of this study is to empirically study the nature of the relation-ship between the prevalence of swing voters and the occurrence of political protests in political systems with two-party competition and a majority or plurality voting system, specifically in the US at the congressional districts level. Accordingly, the question this thesis will seek to answer is:

What effect does the prevalence of swing voters have on the occurrence of political protests?

The question is addressed by using multiple regression analysis to statis-tically examine the relationship between swing voter prevalence and protest occurrence, using data on the concept of elasticity, as developed by jour-nalists at the data journalism site FiveThirtyEight (Silver, 2018), combined with protest data from the Armed Conflict Location & Event Data Project’s US Crisis Monitor (2020). The results of this were, in the end, found to be inconclusive, albeit with a few favorable indications, laying the ground for further studies in the future.

After a review of existing related academic literature, a theoretical frame-work, derived mostly from Rational Choice Theory, is laid out in chapter 2. In chapter 3, the sourcing and processing of the data is presented and discussed, followed by the methodology behind the regression analyses in chapter 4. Lastly, the results of the analyses are presented in chapter 5, and their implications, as well as potential for further research, are discussed in chapter 6.

1.2 Review of earlier literature

In regards to research on polarization, protests and political behavior, there are many potential aspects for research to focus on, including the causes of polarization broadly, the causes and effects of extremist rhetoric, the causes and effects of protests and the factors behind the success or non-success of protest movements. This thesis will focus on the role of polarization as a cause of protests, but research on other and related aspects is prolific and wide-ranging.

(9)

groups in society, by comparing survey data from different countries. They define grievance as individual levels of dissatisfaction with the perceived abil-ity of the political system to provide opportunities to improve one’s standard of living. In contrast to earlier studies, the authors argue that the mean level of grievance in a society is not a good predictor of the occurrence of protests, and points instead to the difference in grievance between different groups. They also find that protests are most prevalent when the average level of grievance in a society is low, but polarization is high. In short, if the level of grievance is high in a substantial minority of citizens, those citizens will resort to protest because the amount of reference groups, i.e., visible groups with better opportunities and standards of living than the aggrieved group, is high (Griffin et al. 2020). By comparison, while also looking at the relation-ship between polarization and protests, this thesis will be using a different measure of polarization and will be examining it on more of a micro-level, comparing congressional districts, a distinct sub-national unit, instead of countries as a whole. The methods and data used here will also allow us to examine the role of grievance as a function of dynamics in the political system, rather than simply grievance per se.

In regards to the persuasion vs. turnout debate discussed earlier, Hall and Thompson (2018) examine how a candidate’s position on the ideological spectrum (moderate to extreme) affects turnout in elections they partici-pate in. Examining a number of close elections between one ideologically moderate and one more extreme candidate, they show that turnout is an important factor for winning elections, but one that generally favors mod-erate candidates over extremist ones. Nominating the latter type tends to lead to increased turnout for the opposing party instead of his or her own, where they instead tend to struggle to increase or even maintain turnout. One possible explanation for this is the concept of concave utility (Hall & Thompson, 2018, p. 512), which will be discussed further later on.

(10)

especially among younger adults. Both of these findings indicate something significant for the purposes of this thesis: what politicians say and do affects citizens’ behavior. While Piazza and Kalmoe examine two separate types of outcomes, domestic terrorism and public support for violence, a contribu-tion here will be to extend the analysis to a third type of outcome, political protests.

(11)

2 Theoretical Framework

The purpose of the following chapter is to present a feasible and theoreti-cally convincing version of how the causal relationship between swing voter prevalence and protest occurrence could be constituted. This theoretical background will also serve as the basis for the guiding hypothesis of this thesis, presented at the end of the chapter.

2.1 Rational Choice

The theoretical background will mostly be derived from Rational Choice The-ory, a theoretical framework commonly used to model and understand human social behavior as a function of the behavior of individual actors. Rational Choice Theory seems suitable here, partly due to its close association with quantitative research methods within the social sciences, and because many of its derivative theories provide a suitable framework for explaining the in-teraction of individual and collective action, specifically within a political system with plurality voting and two-party competition.

A fundamental assumption within Rational Choice is, as the name sug-gests, that individuals act rationally, with the aim of increasing their general wellbeing, or utility. In the words of Downs: “[An individual] proceeds toward its goals with a minimal use of scarce resources and undertakes only those actions for which marginal return exceeds marginal cost.” (Downs, 1957, p. 137). A common misunderstanding of this assumption is that rational action must be the same as objectively good action. For an action to be considered rational, the actor performing it need only believe it to be good. Rationality should be understood as something subjective, in that an individual chooses that course of action which she believes, based on her knowledge and precon-ceptions at the given moment, will bring her the most utility, whether that course of action will objectively do so or not.

2.2 Swing voters

(12)

possible for either candidate to persuade her to vote for them by merit of their chosen policy position alone. A contrasting term which will be used here is base voter, i.e., a voter who, whether they vote at all or not, is more or less completely committed to one party.

In order to understand the variation between a high or low prevalence of swing voters in an electorate and the effect it has on parties’ choice of policy, a good starting point is the Median Voter Theorem, as developed and articulated by Downs (1957). It describes how two-party competition in a system of plurality voting tends to lead to parties and candidates converging on a policy equilibrium in the form of the preferred policy position of the median voter. In Downs’ simple model, all voters in an electorate are lined up on a spectrum from most left-leaning to most right-leaning. Candidates A and B can count on the support of all voters to the left or right respectively of the median voter, and need therefore only convince him or her to win, as shown in Figure 1. Thus, parties will tend to choose policies close to the center.

Figure 1: Voter equilibrium according to the Median Voter Theorem

(13)

between their own preferred policy position and those of the two candidates, always choosing the closest one. (Downs, 1957).

Downs’ theorem in its simplified version can easily appear somewhat sim-plistic or even unrealistic. There are many reasons to think parties’ policy positions can be expected to diverge rather than converge, including observ-able empirical evidence (Stonecash et al., 2003). A more careful reading of Downs as well as derivative and complementary works by other political sci-entists and economists can, however, offer a more nuanced and less absolutist picture. In reality, the theorem should be seen as dependent on a number of influencing factors and assumptions, as described by Downs and others, including Grofman (2004). A number of these factors and assumptions will be discussed next.

2.3 Unidimensionality of issue preferences

As alluded to previously, most rational choice theories about voter behavior, including the Median Voter Theorem, rests on the assumption that voters’ preferences are unidimensional, i.e., that voters care only about one type of issue (most often the foremost being economic issues), and that voters’ preferences can be ordered along a unidimensional scale between two opposite poles (as seen in figure 1). This assumption is relevant here, because an individual’s location on a left-right scale has implications for how prone they are to swing between parties, in that a more extreme policy position (i.e., one far from the center) naturally implies a lower likelihood of changing to the other party.

At a glance, this assumption can seem questionable; most voters obviously care about more than one issue, and different issues can be more or less important for different kinds of voters (an example of a contrasting theory is the so-called GAL-TAN scale by Hooghe et al., (2002)). In practice, however, there is ample evidence that policy positions on different issues often tend to align quite closely, and voters’ and parties’ positions on most issues tend to fit neatly into the same issue dimension. Put differently, left-leaning views on fiscal issues tend to be accompanied by left-leaning views on social issues, and vice versa.

(14)

some issues and left-leaning views on others), is assumed to be a negligible phenomenon. Going forward, voters will therefore be assumed to inhabit a spectrum between two ideologically opposed poles, left and right.

2.4 Distribution of policy preferences

An important way of thinking about the variation in the number of swing voters in an electorate is in the form of different distributions of policy pref-erences (along a spectrum, like previously described) in different electorates, which Downs also points to as an important factor behind convergence (as in the Median Voter Theorem) or divergence of party policy. Figure 2 illus-trates two different electorates: one where policy preferences approximately follow a normal distribution (i.e., where most voters are centered around one “peak” at the center), and one where the electorate is polarized (i.e., where voters are divided in two peaks at opposite ends of the spectrum). Since vot-ers near the center are more likely to switch party, these two electorates can also be seen as one with many and one with few swing voters, respectively.

Figure 2: Policy convergence and divergence in a non-polarized & polarized electorate

In the first scenario, the Median Voter Theorem can be expected to apply (approximately), and the two parties will be pushed to converge, because they gain more votes by moving towards the center. In the second scenario, parties will tend to diverge, for the same reason.

(15)

most voters will also be located relatively close to whichever party is in power, and expect a relatively similar level of utility from either party’s policies. In the second scenario, where voters and parties are more polarized and favor different policies, shifts in power will be associated with more radical shifts in government policy, and one half of the electorate will always feel like policies they are highly opposed to are being imposed on them. The stakes of the political system will, in other words, be higher. If one party is dominant in the latter situation, the supporters of the other party will eventually revolt, while regular shifts in power between the two will result in social chaos, because government policy will shift from one extreme to the other. (Downs, 1957).

Downs therefore concludes that a polarized electorate (i.e., one with few potential swing voters) has a destabilizing effect on society, and inhibits long-term effective and stable government. This conclusion is central to the overall research question of this thesis - whether swing voter prevalence affects the occurrence of protests - because it indicates that fewer swing voters leads to more protests, if one assumes that protests would be a manifestation of such instability.

2.5 Voter Turnout

A further assumption behind the median voter theorem is in regards to the act of voting itself4_{, and is a further factor to whether the two parties’}

poli-cies will converge or diverge. Parties will tend to converge if the only thing affecting whether or not a person actually votes or not are the “costs” of voting, like the distance to a polling location or the opportunity costs of not going to work. If enough voters abstain from voting for other reasons, like neither of the candidates being close enough to their preferred policy positions, fear of losing support among party loyalists can instead push par-ties to start to diverge outwards, towards the respective median or fringe of the respective party’s internal preference distribution, as previously shown in figure 2 (Grofman, 2004 and Lindbeck & Weibull, 1987). As described by

4_{It should be noted that many RC theorists consider the act of voting to always be}

(16)

Lindbeck and Weibull, the difference in expected utility between the policy positions of the two candidates needs to exceed a certain level in order for voters to find it worthwhile to vote at all, when they weigh the expected costs and benefits of going to vote.

Lindbeck and Weibull (1987) as well as Aldrich and McGinnis (1983) also point to the asymmetric importance of loyal party activists compared to av-erage voters, due to their active participation in political campaigns. If these activists, who tend to be more ideologically extreme (Lindbeck and Weibull, 1987), are not sufficiently motivated and decide to abstain from campaign work, the fear of losing their support and active participation can have a diverging effect on candidates’ policy platforms. Their importance would ar-guably increase further, the less persuadable, moderate swing voters inhabit the electorate, further incentivizing policy divergence among candidates in places with a lower swing voter prevalence. Consequently, this further indi-cates that fewer swing voters leads to more extreme candidates.

Additionally, as mentioned earlier, Hall and Thompson (2018) highlight the concept of concave utility as a driving factor behind turnout. If the expected utility of voting for either of two candidates is concave, it means voters will dislike policy more and more the further away from their ideal point they are. That is, a voter’s expected disutility from a candidate’s policy position will increase exponentially the further away it is from his or her preferred ideal point. The result of this, Hall and Thompson find, is that more extreme policy positions taken by candidates tend to be more motivating for voters who dislike them rather than voters who favor them, leading to increased turnout for an extreme candidate’s opposing party (Hall & Thompson, 2018). This speaks against the idea that candidates would be more extreme even in inelastic areas, provided this is something they are aware of and take into account in their strategy (although Hall and Thompsons findings themselves arguably indicates that they do not, since extreme candidates evidently do get nominated).

2.6 Probabilistic voting

(17)

platforms. The theory also holds that while voters are imperfectly informed about candidates’ policy platforms, candidates are also imperfectly informed about voters’ policy preferences, and seek not to strictly maximize their number of votes but rather to maximize the probability to win (Lindbeck & Weibull, 1987).

A derivative model of electoral competition describes a competition be-tween two parties, represented by two candidates, A and B. In this system, the candidates are opportunistic, i.e., mainly motivated by winning elec-tions, and therefore tailor their message in accordance with what they be-lieve will maximize the probability to win in their electoral district (Persson & Tabellini, 2000, p. 47). The model also includes a formula in which four factors decide which of the two candidates an individual voter will vote for: 1 & 2) perceived utility from the policy platform of candidate A and B re-spectively, 3) the voter’s individual ideological bias, or partisanship towards either party or candidate and 4) exogenous shocks (e.g. economic crises) (Persson & Tabellini, 2000, p. 53).

The most significant of those four factors for the purposes of this study is certainly the third, partisanship, since a voter’s partisanship towards either party has strong implications for the likelihood that a candidate from the opposing party would be able to persuade him or her (i.e., whether that voter is a potential swing voter or not). The distinction between swing voters and base voters, can therefore be seen as a distinction between independents and certain weak partisans on the one hand and strong partisans on the other. The main independent variable here, the prevalence of swing voters, can therefore also be defined as the share of a district’s voters made up of partisans or non-partisans, respectively.

2.7 Protests

Borrowing elements from Mancur Olson’s theory of Collective Action (1965), protests will be defined as the public demonstration of a group of individuals aimed at achieving a common objective which they believe will result in increased utility.

(18)

imagine different kinds of political participation along a spectrum from least to most active, on which protesting is more active than voting, and violently protesting more active than peacefully doing so.

(19)

2.8 Analytical framework

To sum up, we can imagine a chain of causal mechanisms. Prevalence of swing voters can be seen as a function of the electorate’s distribution of policy preferences and its aggregate level of partisanship. This prevalence in turn causes convergence or divergence of party policy, which in turn affects individuals’ perceived stakes in the political system. These stakes then finally affect the occurrence of more active modes of political participation, like protests.

Examining this indirect relationship between swing voter prevalence and occurrence of protests is the overall aim of this thesis. This causal relationship is illustrated in Figure 3.

Figure 3: Analytical Framework

(20)

indi-vidual voters. Since voters act rationally, these higher stakes are assumed to motivate voters to participate more actively in the political system, with one (but not necessarily the only) manifestation of this more active participation being a higher occurrence of political protests.

Finally, for reasons explained above, the study will be conducted with the hypothesis that there is a negative relationship between swing voter preva-lence and number of protests, i.e., that fewer swing voters should lead to more political protests:

(21)

3 Operationalization and data collection

The following chapter will describe how the two main variables are oper-ationalized, i.e., how they are practically measured in order to answer the study’s overall question. The data used in the study will also be presented, including how it was sourced and how it was processed before being used in actual analyses.

3.1 Level of analysis

The main level of analysis in the study will be congressional district in the United States. Since there are 435 such districts, 435 will consistently be the number of observations used in analyses.

Congressional districts are the geographical units that form the basis of the composition of the US House of Representatives, the lower chamber of the US Congress. Each state has a number of seats in the House proportional to their share of the total US population (although each state is guaranteed at least one seat). Consequently California, the most populous state, has the most seats, 53, while seven states including Montana, Vermont and Delaware, only have one each. House elections take place every two years, in which each district elects one member separately to a two-year turn, with redrawing and reapportionment of congressional districts occurring every ten years. (United States Census Bureau, 2020).

Considering the theories utilized here, the US seems one of the most obvi-ous choices of country for this study, given its system of plurality voting5 _and

distinct two-party system. Arguably, the US is large and globally significant enough to warrant studying by itself, and the primary goal of the study will be to give an accurate description of the relationship between swing voter prevalence and protest occurrence there. To that end, given that all of the US’ 435 congressional districts are included, the ability of this study to gen-eralize results is of negligible concern, since the study includes all primary units of interest, at least geographically speaking (Teorell & Svensson, 2016, p. 70). That said, to the extent that the US is structurally similar to other countries or units, results could certainly be generalized further, which would no doubt be worthwhile.

5_{While it is true that the elections and systems of voting in the US is administered}

(22)

Nevertheless, the choice of comparing US congressional districts instead of countries or some other geographical unit does come with advantages. First, the US is relatively well-studied compared to many other countries, making it relatively easy to find comprehensive and reliable data, as was the case with the variables used here.

Second, given that congressional districts in the same country exist in the same structural and institutional context, they can be expected to be similar in many ways that, for example, different countries are not. This has several methodological implications in terms of the need to control for potential underlying factors (something that will be discussed more later on), and makes congressional districts a very practical, albeit still varied and worthwhile object of study.

Third, considering that a part of the overall purpose of this thesis is to measure the effect of candidate behavior on voter behavior, using congres-sional districts as a framework is beneficial since candidates would theoret-ically only be incentivized to seek to influence voters in their own district, given that those voters are the only ones in a position to vote for them. Consequently, voters’ expectations and perceptions of the political system can be assumed to derive chiefly from the rhetoric and chosen policy po-sitions of politicians contending in their own district. This, combined with how the structure of the American electoral system itself makes different con-gressional elections truly separate from each other, allows us to theoretically attribute any measured effect of candidate behavior (as a function of swing voter prevalence) on voters in a specific district to that district’s candidates alone.

3.2 Elasticity

(23)

A congressional district’s voter elasticity describes how sensitive its pat-terns of voting are to changes in the national political environment. For example, if a financial crisis leads to a drop in the support for an incumbent president by five percentage points nationally, the drop will be expected to be more than five percentage points in an elastic district, and less in an inelastic one. Consequently, an elastic state or district is more likely to be the subject of bigger shifts in party support, while party support in inelastic states and districts tend to be more stable over time. (Silver, 2012).

To be clear, a district’s elasticity is not the same as its competitiveness, i.e., whether the district as a whole tends to swing between parties from one election to another. Competitive districts (“swing districts”) can be both elastic or inelastic - the difference between the two lies in how large the swings are (Silver & Rakich, 2018). To give an example, a district that shifts between 60% Republican to 60% Democratic in two subsequent elections is more elastic than one that shifts between 51% Republican and 51% Democratic. Both those districts can still be characterized as competitive, however, since both parties seem to have a roughly equal chance at winning in both of them. Moreover, a district or state can be very elastic and be dominated by one party simultaneously, an example of which is the state of Rhode Island. This happens if the state or district has a large share of swing voters and more base voters from one party than the other. For example, imagine a district where 40% of the electorate are Democratic base voters, 20% are Republican base voters and the remaining 40% are swing voters (a relatively high number). In that situation, the number of swing voters needed to reach 50% is lower for the Democrats, and they will therefore win more often. As a consequence, the district will likely tend to shift between mildly and very Democratic, instead of between Democratic and Republican. (Silver & Rakich, 2018).

The district level elasticity scores themselves are, in short, derived from swing voter prevalence6 _{and are meant to be interpreted as the expected}

size of the shift in party support in the district following a shift toward either party by one percentage point nationally (Silver & Rakich, 2018). To

6_{To be more specific, the measurement was developed by analyzing demographic and}

(24)

illustrate, let us view the five most and five least elastic districts, in table 1: Table 1: The five most and least elastic districts

District Elasticity District Elasticity

Michigan 5th _1.24 _{California 2}nd _0.76

Illinois 8th _1.22 _{New York 8}th _0.74

Nevada 4th 1.22 New York 14th 0.73

Massachusetts 1st _1.22 _{Illinois 7}th _0.72

Massachusetts 6th _1.21 _{Pennsylvania 3}r d _0.72

The district with the highest elasticity, Michigan’s 5thdistrict, has a score of 1.24. That means for every one percentage point the support for either party shifts nationally, it is expected to shift 1.24 percentage points in the same direction in Michigan 5th. By contrast, the same shift nationally would only lead to a 0.72 percentage point shift in Pennsylvania’s 3r d _{district, one}

of the nation’s least elastic districts.

3.3 Protests

Occurrence of protests will be operationalized simply as the total number of political protests occurring within the bounds of a specific congressional district.

The protest data used here comes from the US Crisis Monitor (ACLED, 2020), a joint project by the Armed Conflict Location & Event Data Project (ACLED) and Princeton University’s Bridging Divides Initiative (Prince-ton University, 2020). The Monitor aims to collect comprehensive data on political violence and demonstrations in the US, including a wide range of information like dates, locations, fatalities, event types, groups involved and more (ACLED, 2020). ACLED is a US-based non-profit specializing in world-wide conflict data collection, analysis and mapping, and the Bridging Divides Initiative is a Princeton University research project with the aim of building “(. . . ) an expansive picture of peacebuilding and reconciliation organizations across the [US](. . . )” (Princeton University, 2020).

(25)

to its scope; it contained events from every part of the US and included many useful variables, not the least of which was longitude and latitude co-ordinates for each event, i.e., numerical information about its exact location. This geographical information was critical in combining the protest data with the district level elasticity data, as described in the next paragraph. After removing a smaller number of observations due to errors and low-quality ge-ographical information, as well as all events occurring in areas not part of a US congressional district, like the District of Columbia, the final number of events amounted to 153747.

Combining the protest data with the elasticity data into a final, usable dataset was done with the help of the open-source Geographical Information System (GIS) program QGIS (2020). GIS programs generally are frame-works for gathering and analyzing different types of geographical data, like maps (ESRI, 2020). The longitude and latitude coordinates contained in the ACLED dataset were then combined with geographical shapefiles of US congressional districts (GIS-files showing the coordinates of each district) produced by political scientists at UCLA (Lewis et al., 2018), allowing each event to be automatically assigned the corresponding district in which it took place. This then allowed the data to be aggregated to district level, providing a total number of protests for each district.

3.4 Demographic data

The source of the demographic data used for control variables is the 2019 edition of the American Community Survey (ACS) from the United States Census Bureau (2019). The ACS is an annual ongoing survey containing detailed information about demographics, housing, economics and a number of other topics, which is, among others, used by the US government to help determine the distribution of state funds. The population density data is also partially derived from Congressional District Relationship files, also from the

7_{It should be noted that these 15374 events include a number of events that do not}

(26)

US Census Bureau (2017).

The variables used are total population, population density, income, un-employment rate, age, ethnic diversity and education rate. Operationaliza-tions and reasons for including these will be discussed further in section 4.2. Although the data is from 2019 and not 2020 like the protest data, and thus not entirely up-to-date, most variables are assumed to be sufficiently correlated year-to-year to still be fit for use. The only exception is unem-ployment rate, where the 2019 numbers can be assumed to not reliably reflect the situation in 2020, given the exceptionally high unemployment rate result-ing from the COVID-19 pandemic (Bureau of Labor Statistics, 2020). Since the intended purpose of including unemployment rate is as a general and structural indicator of economic conditions in a district, however, using ear-lier data may not only be adequate but also more appropriate, considering what an outlier 2020 is as an economic year.

3.5 Data adjustments

To make analyses and interpretation of results more practical, a few variables will be adjusted or transformed before they are used in analyses.

First, the protest variable will be rescaled from total number of protests in each district to number of protests per capita, a more appropriate unit for comparison. While each district is intended to represent an equal amount of people, mostly for the sake of equal representation, populations vary in reality. An average district represents approximately 700 000 people, but it varies from more than 1 000 000 (Montana’s at-large district) to roughly 530 000 (Rhode Island’s two districts).

Second, the protest per capita variable will also go through a log trans-formation. This means using the natural logarithms8 of the protests’ raw numbers in regression analyses instead of the raw numbers themselves9_{. The}

8_{The natural logarithm of a specific number is the exponent to which the mathematical}

constant e (approximately 2.7) would have to be raised to in order to obtain that specific number.

9_{A minor issue arose with the log transformation, as seven districts had zero recorded}

(27)

regression results of such a log-lin model (a regression model where the depen-dent variable is logarithmic) can be interpreted in terms of relative percentage changes in the dependent variable. For example, a regression coefficient of 0.5 for an independent variable in a log-lin model means that for every in-crease of 1 unit of scale in that independent variable, the dependent variable will increase by 0.5%. Further implications of using a log-lin model will be discussed later on.

(28)

3.6 Descriptive statistics

Table 2 displays some basic information for all variables used in this the-sis. As discussed earlier, the districts with the lowest elasticity scores were Pennsylvania’s 3r d and Illinois’ 7th districts, while Michigan’s 5th and Illi-nois’ 8th_{had among the highest. Meanwhile, seven districts had no recorded}

instances of protests, including California’s 29th _{and New York’s 13}th _and

15th districts, while Oregon’s 3r d district (which includes Portland) had the most, with 225 recorded protests, followed by New York’s 10th _{district, with}

158.

Table 2: Descriptive statistics

Variable Mean St. Dev. Min Max

Elasticity Score 1 0.1 0.72 1.24

Number of protests 35 28 0 225

Protests per 1000 capita 0.05 0.04 0 0.26

Population 752951 59474 529295 1068778 Population Density 939 2668 0.49 28311 Median income ($) 68320 18828 31061 149375 Unemployment rate (%) 4.6 1.4 2.2 10.6 Median age 38.8 3.6 28.7 56 Ethnic diversity (%) 60.2 22.7 2.6 95.3 Education rate (%) 32.7 10.9 10.1 72.5

N = 435. Population density in inhabitants per km2_{. Ethnic diversity is operationalized as}

(29)

4 Methodology

The following chapter will describe the methods used in the study, the pur-pose behind them and certain technical requirements and pre-conditions that need to be met in order to perform them properly.

4.1 Primary method: Regression analysis

Returning to the overall research question of this thesis, whether there is a re-lationship between swing voter prevalence and the occurrence of protests, we can observe that it concerns a causal relationship between two variables, i.e., whether one is the cause of the other. This poses a challenge common to most quantitative research within the social sciences: the objects studied by social scientists, be they countries, individuals or US congressional districts, almost always differ in more factors than the dependent and independent variables of interest, making it hard to reliably rule out any alternative explanation to observed effects. Unknown, underlying factors can easily distort results (Teo-rell & Svensson, 2016, pp. 186-189). If such an underlying variable correlates with the independent variable and somehow affects the dependent variable of interest, it can lead to so-called omitted variable bias. The effect of the omitted variable on the dependent variable is then mistakenly attributed to the included independent variable, showing a false, or spurious relationship (Powner, 2015, p. 203).

In an ideal situation, like much of the natural sciences and some instances in social science, this problem can be dealt with by using an experimental research design, whereby the values of an independent variable of interest are assigned by randomization. In this case, that would have meant randomly assigning an elasticity score to each district and observing the effect it would have on the protest variable. If so, any observed difference in the dependent variable (number of protests) would have been exclusively attributable either to the independent variable (elasticity) or to chance (the probability of which can easily be estimated). In other words, other variables affecting the results in a systematic way can be ruled out, because randomization ensures nothing exogenous affects the distribution of the independent variable10 _(Mendenhall

& Sincich, 2014, p. 170).

10_{In mathematical terms, the expected value of the error term in the regression, given}

(30)

Most often experimental research designs are not feasible, however. That includes this study, since the elasticity variable used here is decidedly not something an individual scientist can have any control over. This leaves us with having to find a way to emulate the ideal situation as best we can. In this study, this will be done by way of multiple regression analysis with the method of ordinary least squares (OLS) and with control variables. OLS-regression fits the purposes of this thesis well, in that it is a commonly used statistical method to measure a statistical relationship between variables (Mendenhall & Sincich, 2014, pp. 167-170). This relationship can also be described mathematically by constructing a regression equation, like so:

Yi = β0+ β1X1,i+ β2X2,i+ · · · + βkXk ,i+ (1)

In the equation above, Yi denotes our dependent variable, β0 denotes the

constant (in short, the value of Y when X = 0); β2X2...k ,i denotes the

re-gression coefficient (the estimated expected change to β0, multiplied by the

value of the independent variable Xk for a given observation i ); denotes an

error term, which can be interpreted as all factors that affect Y apart from X, including chance; and k denotes the number of independent variables in the model.

Control variables are additional factors that are thought to potentially correlate with the primary independent variable and potentially have an effect on the dependent variable of interest. Adding them to a regression means internalizing these other potentially relevant factors into the statisti-cal model, allowing us to isolate the effect of one specific variable, and thus minimizing omitted variable bias. In effect, this means estimating the effect of one independent variable on the dependent variable while all other inde-pendent variables are kept at a constant level (all else equal, or the ceteris paribus assumption, to use a Latin term), thus approximately emulating an experimental research environment (provided all or most potential underly-ing variables can be identified) (Teorell & Svensson, 2016, p. 189). Which control variables will be included here, and why, will be discussed in the next section.

(31)

the other way around. Such reversed causality, i.e., when the directional nature of a relationship is contrary to prior belief, can only be completely ruled out by using other types of research methods, including qualitative ones. In a study like this, however, the theoretical background and analytical framework can serve as sufficient reason to consider it unlikely.

4.2 Control variables

In total, the coming regression analyses will make use of seven separate control variables, namely population, population density, income, unemploy-ment rate, age, ethnic diversity and education rate. These are all thought to be potential underlying variables, meaning every one of them are thought to potentially correlate with the main independent variable of interest (swing voter prevalence) as well as potentially affect the main dependent variable (number of protests)11_.

Population is included simply to account for differences in total popula-tion between districts, as discussed earlier in paragraph 3.5. Since one would naturally expect more protests in an area with more people, and since it is also not unthinkable that population could somehow correlate indirectly with elasticity through other general demographic or political factors (like partisanship), controlling for it seems appropriate.

Population density is likely a highly influencing factor. Lower popula-tion density would plausibly mean fewer people nearby to gather and protest with, as well as having to travel longer distances for such a gathering. From a theoretical perspective, this implies higher costs of such political partici-pation, and would therefore indicate a lower expected number of protests in sparsely populated areas. Since population density is also related to things like geography and occupation, it also likely correlates with certain political characteristics like partisanship, and thus indirectly with elasticity. As seen in table 2, population density also varies significantly in the data.

Unemployment rate and income here serve as indicators of a district’s gen-eral economic conditions, which could plausibly correlate with both our main variables to some extent. First, there are strong reasons to think

unemploy-11_{Consequently, a factor like the impact of the COVID-19 pandemic would not qualify as}

(32)

ment is linked to social unrest (see, for example, ILO, 2013) and, in theory, it would also be a plausible contributing cause of the kind of grievance that, among others, Griffin et al. (2020) has linked to social unrest and protests, hence why controlling for it seems appropriate. Second, since employment status and income likely influence general political behavior, which in turn affects partisanship, it is not hard to imagine that economic conditions corre-late to some degree with elasticity. Income will be operationalized as median household income

Age is included to control for differences among age groups. There is a fairly broad consensus among political scientists and psychologists that younger people tend to participate more in protest and similar types of po-litical activities (for examples, see Watts, 1999; Melo & Stockemer, 2014 and Barroso & Minkin, 2020), attributed to different kinds of lifecycle effects or differences in social and cultural values (Renstr¨om et al., 2020). Since age also seems to correlate with partisanship (possibly for the same two reasons) (Pew Research, 2018), there seems to be enough indications that controlling for it would be appropriate. Age will be operationalized as median age.

Parallel to differences between age groups, there is also good reason to think there are differences between different ethnic groups, making ethnic di-versity a potential underlying variable. A survey by Pew Research found, for example, that both African Americans and Hispanics made up a larger share of participants in protests and rallies for racial equality in 2020 compared to their respective share of the total US population, and vice versa for whites (Barroso & Minkin, 2020). Regarding African Americans, a number of stud-ies (albeit not particularly recent) also seem to support the notion that they engage more in protest than other ethnic groups (see, for examples, Eisinger, 1974 and Secret & Welch, 1982). Ethnic diversity will be operationalized as a district’s share of whites, since white people make up a majority of the population in the US. A higher or lower share of whites is therefore assumed to imply a lower or higher rate of ethnic diversity in a district, respectively.

(33)

bachelor’s degree or higher.

The control variables will be included hierarchically in a total of eight regression analyses, with the eighth and final model including all eight control variables. This model can be expressed as a regression equation like so:

(Log)EventsP erCapitai = β0+ β1ElasticityScore(∗10)i+ β2P opulationi

+β3P opulationDensityi+ β4M edianAgei+ β5U nemploymentRate(%)i

+β6M edianIncomei+ β7EducationRate(%)i+ β8Shareof W hites(%)i+ i

(2)

4.3 Technical regression assumptions

In order to be able to make reliable inferential claims based on the results, certain additional technical assumptions need to be made.

As a first assumption, the expected value of the error term, given any values of the independent variables in the model, needs to be equal to zero12_.

Put differently, omitted variable bias needs to be minimal, as discussed ear-lier. Adding control variables, as previously described, is assumed to have adequately addressed this.

Second, the mathematical equations corresponding to the relationship between the dependent and independent variables in the regression analy-ses are assumed to have the proper functional form. This is addressed by the logarithmic transformation of the protest variable described earlier. The natural logarithm is always smaller than its original number, but bigger num-bers are “pulled down” more than smaller ones, which makes the relationship between the variables approximately linear, and thus suitable when using lin-ear regression. Additionally, log transformations also reduces the potentially distorting impact of high outliers. (Powner, 2016, pp. 185-190).

A third assumption concerns an absence of multicollinearity, i.e., when two or more of the independent variables correlate excessively with one an-other, which can make it difficult for a regression analysis to distinguish the exact source of effects on the dependent variable (Powner, 2016, pp. 190-192). Multicollinearity can be detected by performing a so-called variance inflation factor (VIF) test on a regression model, which mathematically pro-vides a measurement of the presence of multicollinearity in any independent

12_E(|X

(34)

variable (Pennsylvania State University, 2018). Such tests will be performed here13.

The last assumption concerns homoscedasticity or constant variance, mean-ing that the variation in the dependent variable needs to be approximately the same across the range of values of each of the independent variables (Mendenhall & Sincich, 2014, p. 170). Even if there are no particularly evi-dent reasons to think this will be the case here, a simple way that will be used to hedge against any possible such issues is to use heteroscedasticity-robust standard errors, which is a method to produce unbiased standard errors un-der potential heteroscedasticity (University of Virginia, 2020).

4.4 Interpretation of results

A final point before moving on to results concerns statistical significance. The past few years has seen a debate over what many see as the improper or excessive use of statistical significance and significance testing across many scientific disciplines, which critics argue is based on a misguided and arbi-trary notion that scientific results can always be dichotomously divided into significant and non-significant. Such critics instead emphasize that statistical results need to be viewed in a more nuanced way, and, as an example, sug-gest focusing more on showing and discussing confidence intervals. (Nordin, 2020; Amrheim et al., 2019).

While this debate will not be further elaborated on here, such sugges-tions, as well as the discussion as a whole, will be taken into account in the later parts of this thesis, mainly through a focus on confidence intervals and compatibility assessments (estimating what effects are most compatible with the results of the study) rather than whether results are dichotomously significant or not.

13_{Multicollinearity can also be detected, albeit less reliably, by looking at the correlation}

(35)

5 Results

Table 3 displays eight regression analyses hierarchically, in which additional control variables were added to each consecutive model, approximately in order of importance14_.

Table 3: Regression Results

Dependent variable: (Log) Events per capita

(1) (2) (3) (4) (5) (6) (7) (8) Elasticity Score (*10) .157 .146 −.046 −.052 −.058 −.053 −.027 −.042 (.094) (.095) (.067) (.068) (.068) (.068) (.070) (.072) Population −.267 −.341 −.319 −.344 −.352 −.400 −.365 (.093) (.087) (.089) (.092) (.094) (.090) (.092) Population density −.0002 −.0002 −.0002 −.0002 −.0002 −.0002 (.0001) (.0001) (.0001) (.0001) (.0001) (.0001) Median age .017 .012 .012 .008 −.013 (.017) (.016) (.016) (.016) (.014) Unemployment rate (%) −.053 −.037 .004 .064 (.046) (.046) (.045) (.054) Median income .003 −.015 −.011 (.003) (.005) (.006) Education rate (%) .041 .037 (.011) (.012) Share of Whites (%) .009 (.005) Constant −11.95 −9.82 −7.14 −7.91 −7.22 −7.45 −7.53 −7.76 (0.96) (1.39) (0.99) (1.12) (1.15) (1.18) (1.13) (1.14) Observations 435 435 435 435 435 435 435 435 Adjusted R2 _0.01 _0.02 _0.22 _0.22 _0.22 _0.22 _0.25 _0.26

Heteroscedasticity-robust standard errors in parentheses. Population in 100000s. Population density in inhabitants per km2_{. Median income in thousands of US dollars. Keeping in mind the earlier discussion}

about statistical significance, significance stars are intentionally left out.

Viewing the results, the regression coefficient for the elasticity variable (on the first row of numbers) varies somewhat, depending on which control vari-ables are present. The population density variable seems to have the most notable impact, which becomes apparent when comparing the elasticity

co-14_{After performing the regressions, a VIF-test was performed on each model (except for}

(36)

efficients in models 2 (0.146) and 3 (-0.046) (as well as the adjusted R2_).

Controlling for population density not only significantly changes the effect of the elasticity variable, it also changes from positive to negative. Controlling for age, unemployment rate and income seems to have relatively marginal impact, while controlling for education rate and racial diversity is somewhat more impactful.

All eight models are included for the sake of transparency, but the most important one is certainly the eighth, where all seven control variables are included. In Model 8, the regression coefficient of the elasticity variable is -0.042, which can be interpreted as an expected decrease in the number of protests per capita by roughly -4.2% for every 0.1 point increase on the elasticity score.

Reiterating the hypothesis, that higher elasticity (i.e., more swing voters) leads to fewer protests, these results seem favorable at a first glance. The model’s best guess, an effect of -0.042, is, in fact, relatively consequential. In other words, a higher elasticity score (i.e., a higher prevalence of swing voters) do seem to cause a reduction in the number of protests per capita in the district, as hypothesized.

(37)

Table 4: Confidence intervals (95%) of regression coefficients for the elasticity variable in each regression model, 1-8

Model Lower bound Estimate Upper bound CI Range SE

(1) -.027 .157 .342 .370 .094 (2) -.041 .146 .333 .373 .095 (3) -.177 -.046 .086 .263 .067 (4) -.186 -.052 .081 .267 .068 (5) -.192 -.058 .075 .267 .068 (6) -.187 -.053 .080 .267 .068 (7) -.164 -.027 .111 .275 .070 (8) -.184 -.042 .099 .283 .072

Dependent variable: (Log) Protests per capita. Independent variable: Elasticity Scores (*10). Heteroscedas-ticity robust SEs.

As seen in the interval of model 8 on the last row of the table, the possible effects that are most compatible with our results range from –0.184 (18.4% protest per capita decrease for every 0.1 point elasticity score increase) to 0.099 (9.9% protest per capita increase for every 0.1 point elasticity score increase). Because the interval covers zero and thus contains both negative and positive values, it is difficult to reliably rule out a positive effect, which would be contrary to the hypothesis. However, since a larger part of the interval covers negative values than positive ones, a negative effect seems more likely than a positive one, which is favorable to the hypothesis.

(38)

5.1 Sensitivity analysis

A way of assessing the robustness of statistical results is to perform a so-called sensitivity analysis, which means analyzing how results stand up to various kinds of changes or tweaks to the underlying data, or when (tentatively) addressing potential sources of errors. This could include removing potential outliers or changing values to reflect hypothetical scenarios. Given the limited scope of this thesis, two such analyses will be performed here.

A first potential problem concerns possible interactive effects between districts and their surroundings. The elasticity measurement is constructed by looking at the residents of a district (Silver, 2012), but a protest in a given district is not necessarily made up exclusively of participants residing in that district – individual participants could always have come from else-where. Consequently, districts may not be completely independent of each other, which can lead to bias (Cameron & Miller, 2015). This could also impact the validity of the study negatively, if the effects of a district’s elas-ticity on its people’s tendency to protest could manifest itself in an adjacent district, and thus make certain protest numbers misleading. One potential method of remedying this is to use cluster robust standard errors in a regres-sion analysis, which take into account correlation within certain clusters of observations when clustering them together according to a certain variable. In this case, the observations will be clustered according to state, because of differences in population density between them. The results, which can be viewed in detail in the appendix, show no improvement, however. To the extent that this might be because of methodological flaws, testing this using more sophisticated methods may be a worthwhile subject of future studies.

(39)

a separate regression analysis was performed on each group15_{. The results}

of these analyses, which are also available in the appendix, indicate that the effect is strongly negative in Democratic-leaning districts, moderately nega-tive in toss-up districts and slightly posinega-tive in Republican-leaning districts. In other words, partisan lean to the left or right seems to amplify or suppress the effect of elasticity, respectively. However, given that the results exhibited even more uncertainty than before, further studies of the role of partisan lean would probably be beneficial.

15_{The groups were based on the districts’ partisan lean as reported by the Cook Partisan}

(40)

6 Conclusion & Discussion

Let us return to the question posed in the beginning of this thesis: What effect does the prevalence of swing voters have on the occurrence of political protests? Given the high levels of uncertainty exhibited by the results of this study, any conclusions should be considered tentative. The best estimate, a negative effect of -0.042, is a good indication that the hypothesis, that fewer swing voters leads to more political protests, is correct, and that the theoretical framework laid out in chapter 2 seems to have held. To the extent that they can be considered reliable, then, the empirical results of this study seem to support the theoretical argument made by Downs (1957), that a polarized electorate inhibits long term stable and effective government and has a destabilizing effect on society (the American society, in this case), as discussed back in chapter 2.

In order to be able to draw less uncertain conclusions, further research is certainly warranted. The remainder of this thesis will be spent discussing how future studies can improve and expand on the results of this one.

One of the most obvious ways to reduce uncertainty in regression results is to simply increase the size of one’s sample. Having many observations can reduce the potential impact of outliers as well as generally reducing the impact of idiosyncrasy or chance. While 435 may not be a low number per se, it is absolutely feasible to include more data in a similar study. This could include protest data from a significantly larger time frame than only 6 months - ideally from multiple years - and an updated and temporally differentiated elasticity variable that could reflect changes in elasticity in districts over time, if possible. The availability of such data is certainly feasible, although not guaranteed.

In terms of generalizing the results of this study to an international con-text, the results shown here are assumed to also apply to other countries to the extent that the US is comparable to them. To the extent that the US is not, studying the relationship between swing voter prevalence and protest oc-currence comparatively between countries with, for example, different kinds of electoral systems or systems of party competition would certainly be a worthwhile aim of a future study. In his article about the Median Voter Theorem, Downs (1957) includes a discussion about political systems with proportional voting and multi-party competition, which could, for example, serve as part of the theoretical basis for such a comparative study.

(41)

think about the electoral characteristics of political units in democracies. Be-yond, for example, merely looking at competitiveness, i.e., whether a district or state tends to shift between parties, knowing the size of those shifts adds a layer of complexity that can give us many clues as to how an individual state or district behaves electorally, and by extension how it can be expected to respond to certain types of campaigning or other aspects of the democratic process.

While the potential of the elasticity concept within political science broadly seems large, the specific measurement used here remains virtually untested in a scientific context. Naturally, the ability of the elasticity scores to accu-rately predict voter shifts is dependent on the correct assumptions and sound methodology of its developers, as with any measurement. While there are no particularly apparent reasons to doubt the methodology16 _{of Silver and}

Rakich (2018), it is difficult to determine the robustness of the measurement based on this study alone, wherefore it would be worthwhile to use it again in future studies. Additionally, using other methods to calculate the size of voter shifts could also serve to validate the measurement’s findings, an ex-ample of which could be using actual election results, akin to the methods used by the Cook Political Report to calculate districts’ and states’ partisan lean (The Cook Political Report, 2018).

While the robustness of protest data such as the ACLED data used here seems relatively solid, additional or unused variables or aspects of the data could certainly be the subject of future study. For instance, this study does not account for what kinds of protests occur in specific districts (i.e., whether a district has more protests associated with left-leaning or right-leaning causes), and whether it differs systematically between districts. Studying this could provide insight into how different groups, ideological or otherwise, use protests at different rates and, by extension, under what conditions dif-ferent actors resort to protesting.

Furthermore, let us turn to the validity of the study, which can be de-scribed as whether the operationalized indicators of our central concepts are consistent with their corresponding theoretical definitions — in other words, if the indicators actually measure what they are intended to measure (Teo-rell & Svensson, 2016). The most apparent potential validity problem has

16_{To reiterate, the measurement was developed by looking at demographic and political}

(42)

already been addressed and discussed, namely the issue outlined in section 5.1, earlier, concerning potential interdependence between districts and their surroundings. Apart from that, no particularly notable validity issues seem present.

In terms of reliability, which can be explained as the absence of any non-systematic measurement errors, and the ability of others to perform the same analysis and receive the same results (Teorell & Svensson, 2016, p. 56), any issues are assumed to be negligible. The statistical methods used here are all part of established practice, and are assumed to have been performed correctly17_{. Additionally, the interpretation of data and results, assuming it}

is done correctly, should not differ from one person to another.

Finally, let us reiterate the hypothesis of this study, that fewer swing voters (i.e., lower elasticity) leads to more protests. The way this study is structured, it only aims to tentatively identify the size and nature of that relationship (the if ) — it does not concern itself with its causal mecha-nisms (the how ), or directional nature. While a theoretical rationalization for that relationship as well as a possible causal chain is included here, this study is more empirically than theoretically oriented. A closer study of that causal chain would, among other things, allow us to rule out potential re-versed causality, as discussed earlier. Other possible methods for such studies include using regression analysis with potential mediating variables, which could give a clearer picture of the causal chain. Individual level surveys or qualitative methods like interviews would also be good complements to this study, and could for example include interviewing political candidates about how they perceive things like elasticity, as well as how they respond to it.

17_{All data processing and analyses were performed in the programming language R.}

(43)

7 References

Abramowitz, Alan I (2010). The Disappearing Center: Engaged Citizens, Polarization, and American Democracy. New Haven: Yale University Press.

Aldrich, John H & McGinnis, Michael D (1983). A model of party constraints on optimal candidate positions. Mathematical and Computer Modelling 12(4): pp. 437-450.

Amrheim, Valentin; Greenland, Sander; McShane, Blake (2019). Retire Sta-tistical Significance. Nature 567(7748): pp. 305–307.

Armed Conflict Location and Event Data Project (ACLED) (2020a). About ACLED. https://acleddata.com/about-acled/#1597711250928-21e0636 2-0fb3 (Accessed: 15 November 2020)

ACLED (2020b). US Crisis Monitor. https://acleddata.com/special-proj ects/us-crisis-monitor/#1594244155121-20015e6c-43d0 (Accessed: 15 November 2020)

ACLED (2020c). Demonstrations & Political Violence in America: New Data for Summer 2020. https://acleddata.com/2020/09/03/demon strations-political-violence-in-america-new-data-f or-summer-2020/ (Accessed: 15 November 2020)

ACLED (2019). Armed Conflict Location & Event Data Project (ACLED) Codebook. https://acleddata.com/acleddatanew/wpcontent/uploads/d lm uploads/2019/04/ACLED Codebook 2019FINAL pbl.pdf (Accessed: 15 November 2020)

Barroso, Amanda & Minkin, Rachel (2020). Recent protest attendees are more racially and ethnically diverse, younger than Americans overall. Pew Research. https://www.pewresearch.org/fact-tank/2020/06/24/rec ent-protest-attendees-are-more-racially-and-ethnically-diverse-younger -than-americans-overall/ (Accessed: 16 November 2020)

BBC (2020). Coronavirus: US protests against and for lockdown restrictions. BBC News, US & Canada. https://www.bbc.com/news/av/world-us-c anada-52344540 (Accessed: 5 November 2020)

Brady, Henry E; Verba, Sidney; Lehman Schlozman, Kay (1995). Beyond Ses: A Resource Model of Political Participation. The American Political Science Review 89(2): pp. 271-294

(44)

Cameron, A Colin & Miller, Douglas L (2015). A Practitioner’s Guide to Cluster-Robust Inference. Journal of Human Resources 50(2): pp. 317-372.

Cook Political Report (2018). PVI. https://cookpolitical.com/pvi-0 (Ac-cessed: 17 December 2020)

Cook Political Report (2017). Introducing the 2017 Cook Political Report Partisan Voter Index. https://cookpolitical.com/introducing-2017-cook-political-report-partisan-voter-index (Accessed: 17 December 2020) Croissant, Yves; Millo, Giovanni (2008). Panel Data Econometrics in R: The

plm Package. Journal of Statistical Software 27(2): pp. 1-43.

DeFord, Daryl; Duchin, Moon; Solomon, Justin (2020). A Computational Approach to Measuring Vote Elasticity and Competitiveness. Statistics and Public Policy 7(1): pp. 69–86.

Dimock, Michael & Wike, Richard (2020). America is exceptional in the nature of its political divide. Pew Research. https://www.pewresearch. org/fact-tank/2020/11/13/america-is-exceptional-in-the-nature-of-its-p olitical-divide/ (Accessed: 23 November 2020)

Downs, Anthony (1957). An Economic Theory of Political Action in a Democracy. Journal of Political Economy 65(2): pp. 135-150.

Eisinger, Peter K (1974). Racial Differences in Protest Participation. The American Political Science Review 68(2): pp. 592-606.

ESRI (2020). What is GIS? https://www.esri.com/en-us/what-is-gis/overv iew (Accessed: 17 December 2020)

John Fox and Sanford Weisberg (2019). An R Companion to Applied Re-gression. Third Edition. Thousand Oaks: Sage

Graham, Nathaniel; Arai, Mahmood; Hagstr¨omer, Bj¨orn (2016). multiwayv-cov: Multi-Way Standard Error Clustering. R package version 1.2.3. https://CRAN.R-project.org/package=multiwayvcov

Griffin, John D; Kiewiet de Jonge, Chad; Velasco-Guachalla, Vania Ximena (2020). Deprivation in the Midst of Plenty: Citizen Polarization and Political Protest. British Journal of Political Science: pp. 1-17.

Grofman, Bernard (2004). Downs and Two-Party Convergence. Annual Review of Political Science 7(1): pp. 25-46.

Hall, Andrew B & Thompson, Daniel F (2018). Who Punishes Extremist Nominees? Candidate Ideology and Turning Out the Base in US Elec-tions. The American Political Science Review 112(3): pp. 509-524. Hlavac, Marek (2018). stargazer: Well-Formatted Regression and Summary

Voter Elasticity and Political Protest A quantitative study in an American context