Why Intervene?

Scholars have long known that democratic elections matter; such contests have mostly been treated as an independent variable. Research on elections as a dependent variable – that is, an event that attracts international attention and action – has focused primarily on the role of external actors as promoters of democracy. In pioneering work, Hyde (Reference Hyde2011) has shown why and how states invite external election observers. In many cases, a trade-off is seen between fighting for freedom and fairness, and getting the right person to win the elections (Donno Reference Donno2013; Kavakli and Kuhn Reference Kavakli and Kuhn2020).Footnote ² A number of studies have examined the issue of taking sides – foreigners intervening in a partisan manner in elections. Corstange and Marinov (Reference Corstange and Marinov2012) and Shulman and Bloom (Reference Shulman and Bloom2012) show how voters evaluate partisan foreign interventions in their elections. Levin (Reference Levin2016) argues that superpower support enabled targeted candidates to win during the Cold War. Tomz and Weeks (Reference Tomz and Weeks2019) investigate how Americans would like to push back against hypothetical intervention scenarios in their elections. We extend the insights from these studies by assessing which factors provoke interventions. We focus on what we believe is an important factor – incumbent–opposition policy divisions.

Hypotheses

We set up a very simple decision-theoretic model. Take an election in which two candidates propose policy platforms that may be different from the point of view of the outside power. We assume the incumbent (A) and opposition (B) parties have policy platforms denoted by A _gov and B _opp. An incumbent's vote share can be represented as:

$$f( {{\rm voter}\,{\rm support, \;}\,{\rm election}\,{\rm bias}} ) = f( {\chi , \;\beta } ) $$

We assume, in keeping with observed empirical regularities, that a more biased election (rigged vote count, difficult registration, etc.) benefits the incumbent's chances of re-election, whereas a cleaner election benefits the opposition. We also assume that more votes for one party will benefit that party and may lead to its victory. The incumbent wins, and the opposition loses, through a combination of rigging the rules and garnering votes. Either may work by itself, and the success of one may partly depend on the other (fewer votes may result in a higher final tally under more biased rules).

We further assume that an intervener has two options: (1) support candidates directly, by investing resources in their campaign (for example, promising aid if they win) or (2) alter the election rules, for instance by insisting on a cleaner election or tolerating repression of opposition activists. We refer to the former as candidate or c investments, and to the latter as process or p investments. The two distinct interventions change candidate support: χ(c), and the bias: β(p). The post-intervention vote share therefore becomes: f(χ(c), β(p)) = f(c, p).

We assume that the election result determines the policies according to the vote shares of the platforms A _gov and B _opp and that interventions are costly.Footnote ³ Depending on the degree Γ to which the outsider cares about the result of the election, the intervener's utility will be:

$$u( {c, \;p} ) = \Gamma ( {\,f( {c, \;p} ) \cdot A_{{\rm gov}} + ( {1-f( {c, \;p} ) } ) \cdot B_{{\rm opp}}} ) -s( {c, \;p} ) $$

The problem of intervention then becomes how many resources to invest in an election in order to bring policies closer to the more desirable platform, given one's preferences for the policies of one candidate over the other.Footnote ⁴

For a more compact statement, we define polarization as: π = A _gov − B _opp, such that the intervener prefers the government to win when π > 0 and the opposition to win if π < 0. Polarization refers to the difference in the potential intervener's utility of the incumbent or opposition winning (re-)election. Then the outsiders’ choice is to pick c and p to maximize:

$$\mathop {\max }\limits_{c, p} \quad \pi \Gamma f( {c, \;p} ) -s( {c, \;p} ) $$

We assume that the supported candidate (with c) is more likely to win, and that more bias means the incumbent wins. If we further denote incumbent-supporting interventions as c > 0 (with c < 0, opposition support), and pro-democracy or bias-reducing interventions as p > 0 (p < 0, democracy-eroding, or bias-increasing ones),Footnote ⁵ we can make the following predictions:

We formally derive these hypotheses for a suitable choice of functional forms in Appendix A.Footnote ⁶ The first hypothesis simply says that if an outsider wants a candidate to win, they will invest in that candidate's vote share. The second hypothesis predicts that outsiders can also – in addition or instead – invest in influencing the rules in a manner that favors their candidate. In different cases, the mix of process and candidate interventions would vary depending on local conditions and intervener preferences. What concerns us is the sign of the slope of the investment in either strategy over polarization, which we argue will hold.

We would like to emphasize that the difference between the government's and the opposition's policy positions is key to interventions. When the polarization parameter π is set to 0, interventions of either kind only generate costs and would thus not be undertaken.

We also note an additional implication of this model. We capture importance with the parameter Γ. Policy divisions on more important issues (higher importance Γ) will attract interventions. Higher Γ increases the (absolute value) of the predicted slope of interventions. The issue at stake is expected to matter in motivating interveners to take an interest in an election.

Our model can easily accommodate differences in the cost of intervention across countries. For example, we would expect the cost of process interventions to be much higher in consolidated democracies (and hence, they would seldom be undertaken). The reason is that working institutions make it more difficult to affect how votes are counted, for instance. Note that candidate interventions remain a possibility: voters in consolidated democracies and elsewhere can be swayed by foreign campaign funds and similar measures.Footnote ⁷

We assume that foreign help can boost the fortunes of a domestic ticket. As long as that works at least sometimes, our model applies. We know that in many cases domestic actors ask for foreign support and claiming it is key to their success – for instance, in the case of Greek post-war elections up until the 1967 military coup.Footnote ⁸ If foreign powers assume that interventions will backfire, they may refrain from undertaking them, or they may be more likely to intervene in a covert manner.

The Role of Democratic Outside Power as a Driver of Interventions

So far, we have assumed that the intervener has no inherent preference for cleaner rules, and would invest whatever maximizes the return on policy in light of the cost. According to our model, even non-democracies would sometimes promote democratic elections if this would enhance the effectiveness of pro-candidate interventions. For example, in the 1945 Finnish election, the Soviet Union demanded that the communists should be given more freedom to campaign, and more access to the media. The Soviets based their demands on threats to Finish post-war recovery and territorial integrity.Footnote ¹⁰

We add an additional term to the maximization problem, to give the intervener an externality from investing in cleaner rules. Democracy promotion may carry an additional positive externality for a pro-democracy power:

$$\mathop {\max }\limits_{c, p} \quad \pi \Gamma f( {c, \;p} ) -\lambda \beta ( p ) -s( {c, \;p} ) $$

The additional parameter λ > 0 captures the extent to which the foreign power cares about democracy – more precisely, the (non-negative) level of bias in the conduct of an election. Process interventions may be justified either because a power cares about more or less democracy, understood as cleaner elections, or (in addition to or solely) because candidate interventions are enhanced or stymied by changing the rules of the game. Another way to state this prediction is: in periods when democracy matters more for states, more pro-democracy interventions will be undertaken. In the next section we demonstrate how a novel data set helps to demonstrate the utility of this theoretical exercise.

The Process–Party Dataset

We start by identifying all (potentially) competitive global elections over the period 1945–2012 using the National Elections Across Democracy and Autocracy (NELDA) dataset (Hyde and Marinov Reference Hyde and Marinov2012). Following Hyde and Marinov (Reference Hyde and Marinov2012), we require that at least the following minimal conditions for political competition are present: more than one candidate or party competes, and the incumbent's office is subject to electoral competition (these are variables available in NELDA). All regimes we study are ‘shades of democracy’, though some may be more on the authoritarian spectrum. We then take a stratified random sample: for each country, we pick one Cold War and one post-Cold War election (the period may be important for patterns of interventions). We rely on random selection to produce results representative of the general relationship between polarization and interventions.

In a next step, for all elections in our sample, we predefined a list of powers that potentially had a stake in the contest. We rely on the following selection criteria: all global/great powers (veto-yielding members of the UN) and regional powers, former colonial powers, immediate neighbors and regional political organizations. Potential interveners are selected regardless of whether they actually intervened. The full list is part of the online codebook. Each election has eight potential interveners on average.

Our next objective is to use a variety of sources to generate data for each election on potential and actual interveners, candidate interventions, process interventions, and the stances of the government and the opposition toward the foreign intervener. We record the policy stances of the local actors in a manner that precedes and is independent of any actual intervention.

To find appropriate information, we asked research assistants to conduct targeted searches for each election cycle of databases, such as LexisNexis, ProQuest, Google News, historical newspaper archives, policy content generated by the actors themselves, and scholarly work related to a particular election in journals and books. This procedure ensured access to a variety of English-language sources from around the world. In addition, the researchers processed French, German, and Spanish sources. The resulting articles were checked for appropriate information on interveners, policy stances and relations. Once we located information, we saved it and provided coding notes to document the coding choices.Footnote ¹¹

Candidate interventions try to influence the number of votes cast for a particular candidate, often (but not necessarily) within the legal framework of an electoral system. Candidate interventions include a state or organization that (1) makes benefits, such as foreign aid, conditional on who wins, (2) threatens to pursue negative consequences (economic sanctions, suspending diplomatic relations, military threat) in response to the ‘wrong’ winner, (3) provides monetary or other benefits to a specific candidate or party, or (4) makes statements in support of the election of a specific candidate or party. This includes public criticism of actual or proposed policies if the criticism can be linked to support for the government or opposition at the ballot box.

The full set of allowed candidate codings includes: 1 = for support for incumbent/government candidate; 0 = no support for any candidate; –1 = support for an opposition candidate. The intermediate categories of 0.5 and –0.5 refer to actions that are less clear, or that conflicting information is present, with some evidence indicating support for the opposition but other (less significant) evidence contradicting that.

The most frequent type of candidate intervention in the dataset is public endorsements (32 per cent). A full 25 per cent of candidate interventions qualify as providing support for a party, such as campaign funds or propaganda. Promises of economic rewards or club membership (for example, in the European Union) that are tied to a candidate make up 16 per cent of candidate interventions. Only about 1 per cent of candidate interventions are candidate-specific military threats or acts of violence. Most interventions include more than one action, such as linking aid to a candidate's election and joint campaign appearances with the US ambassador.

Process interventions capture whether a foreign power sought to enhance ( + 1), weaken ( − 1), or exerted no effect (0) on the electoral process. Values of 0.5 and −0.5 are allowed for less clear (mixed) interventions, with one type dominating. Some examples of a positive process intervention include a state or organization: (a) sending or supporting credible election monitors to testify about the fairness of an election, (b) attempting to even the playing field in accordance with national electoral laws and internationally accepted standards, (c) ensuring orderly and free and fair elections by providing security for voters, or (d) threatening consequences if fraud occurs, such as economic sanctions or freezing aid. Examples of negative process interventions include if a state or organization: (1) tries to legitimize an election process that is clearly corrupt, (2) helps governments change voting rules in a way that distorts the playing field, (3) endangers orderly and free and fair elections by approving or assisting in violence against voters or a political group or (4) provides benefits to governments that orchestrated electoral fraud.

Again, noting that most interventions feature more than one element, we observed predominant patterns. Praising polls with widely recognized shortcomings (31 per cent) was the most frequent type of negative intervention, followed by playing down evidence of bad play (23 per cent) and actively supplanting democratic institutions, which occurred in 11.4 per cent of all negative process interventions. The following positive interventions were the most common: observing elections (35.3 per cent), providing positive reports and rewards for clean elections (31.3 per cent), providing negative reports and punishments for flawed elections (22.3 per cent), and providing aid, security or logistics for the electoral process (12.7 per cent).

All actions coded in either the candidate or process support variables concern only the pre-electoral period. We seek to avoid judgement about the effectiveness of different intervention strategies, so we merely record whether an intervention of any kind has occurred.Footnote ¹²

According to our argument, policy polarization accounts for outsiders’ decisions about whether to intervene in an election. The variables A _gov and B _opp capture the orientation of the government and opposition toward the intervening power. More precisely, this is the stance of the relevant actor on the policy issues a foreign power cares about in relation to that country. Government is defined as the political party or parties and/or leader in office at the time of the election.Footnote ¹³ The opposition is defined as the opposition party or parties in parliament, their frontrunners, and/or new parties or candidates attracting a substantial amount of public support.

A _gov and B _opp range from friendly ( + 1) to neither friendly nor unfriendly (0) to unfriendly (− 1). Intermediate values of 0.5 and −0.5 denote mixed messages on relations with the foreign power, with one direction predominating. While we consider and document stances on all issues of concern to the outside power in an election, the orientation coding is a single numeric value that reflects the human coder's judgment on the overall stance of the actor toward the intervening power. If there is more than one government or opposition party, the coding takes the strength of each party in previous elections or in the pre-election polls into account.Footnote ¹⁴

Again, we clarify that we focus on the pre-election period: no information from the period after the election influences our coding decisions. Information from scholarly books and articles is only considered if it pertains to the pre-election period. The publication date of scholarly books or articles does not need to be before the election date, but the information provided by these sources must be. If we were unable to identify whether they apply to the pre-election period, they were not considered for the coding decisions.

The measurement for polarization π is based on the difference between A _gov and B _opp. Consequently, π ranges, in increments of 0.5, from +2 (friendly government, unfriendly opposition) to –2 (unfriendly incumbent, friendly opposition).Footnote ¹⁵

Table 1 shows the coding scheme at work in nine elections. The data collection takes as its starting point the NELDA dataset (Columns 1 and 2). Columns 3–7 come from our new coding. In the Brazil-as-intervener case in the Uruguayan 2004 legislative election (165-2004-1031-L1), the governing party in Uruguay had neither a friendly nor unfriendly attitude toward Brazil (0), and the opposition parties had a friendly stance toward Brazil. According to the last two columns, Brazil supported the opposition with a candidate-targeting election intervention but refrained from seeking to affect the rules of the game.

Table 1. Illustrative codings

We documented the grounds for each decision. There are case studies for each intervener in each election, in which information from the original sources is saved. Our method of annotating the case studies allowed us to use natural language processing text analysis tools. We highlighted portions of the original text with pre-defined ‘keys’ that justify a specific coding of a variable. We implemented the highlights as a custom-created simple mark-up language in LATEX that is rendered as colors in the resulting PDF file for the convenience of coders and readers; the choice of colors is arbitrary. In addition to transparency, this type of documentation offers other advantages, which we discuss in the next section.

Automatic Identification of Directionality

We want to know whether a machine-learning algorithm would be able to produce numeric values from the saved, annotated notes that are similar to the ones our human coders provide. In particular, we seek to find out whether ‘positive’ and ‘negative’ directions in government and opposition classifications are so consistently coded that an algorithm could learn to precisely distinguish them. To that end, we retrieve all snippets with their related code. In this setting, each highlighted piece of text is associated with a binary label (− or +), referring to either an overall negative or positive stance toward the foreign country. We employ a Support Vector Machine (SVM), a very popular supervised-learning model for classification, adopted in both natural language processing research (Joachims Reference Joachims1998) and the text-as-data community (D'Orazio et al. Reference D'Orazio2014; Merz, Regel and Lewandowski Reference Merz, Regel and Lewandowski2016).

It is crucial for supervised-learning approaches to provide training examples, which are pieces of text that either fit a certain description or not (Jäger Reference Jäger2018, 509). The SVM classifier uses the example to map the text as points in a multi-dimensional space. Subsequently, it finds the vector that separates points belonging to the two classes (in our case, these are highlighted strings that point to negative or positive stances by the government or opposition). Then, when the SVM is provided with a new snippet of text that needs to be classified, it maps it again into that same space and predicts the class, based on the position of the point with respect to the separating vector. Based on this distance, the SVM can also provide us with the level of confidence of the prediction.

We test the SVM in a tenfold cross-validation setting (see Manning, Manning and Schütze Reference Manning, Manning and Schütze1999) in comparison with a different classification algorithm, a Naive Bayes (NB) classifier. For each, we experiment with both term-frequency (TF-IDF) and semantic (w-emb) vector representations of texts.Footnote ¹⁶ We are able to achieve a good performance with the different classifiers – up to 80 per cent accuracy with the SVM employing word embeddings – which means that the property under study is better captured when modelling the underlying semantics and not just the frequency of words.Footnote ¹⁷

We used this algorithm in a cross-validation setting to produce machine-generated predictions for all highlighted texts in our collection.Footnote ¹⁸ The SVM classifier also gives us the opportunity to move beyond the binary label provided for each highlighted text. In order to obtain a continuous value, ranging between −1 (highly negative) to 1 (highly positive), we employ the confidence score of each decision and re-scale each prediction.

This provides us with a new version of the dataset, in which machine-generated continuous values replace human numeric decisions (−1, −0.5, 0, 0.5, 1). The continuity in the measurement reflects the computer learning from the general direction assigned by the human coder, which extrapolates and assigns some uncertainty to its decision. This gives us an alternative measurement of polarization, which we use to check the robustness of our results.

Automatic Identification of Policy issues at Stake

International relations tends to divide issues of concern in relations between states into broadly economic and military/political categories. Past studies have generally treated military-related issues of security as dominant since they pertain to state survival (Gilpin Reference Gilpin1987).

We may want to know, for instance, whether the discussion of relations between Bolivia and the United States centers on debt restructuring, armed conflict or drug production. A Bolivian politician may either promise to eradicate drugs (a positive development from a US point of view) or the opposite (a negative development). For convenience, we refer to the stances as positive or negative (according to the directionality discussed in the previous section), and to the policy issues as those at stake in an election. A policy issue may be salient, but it may not generate divisions if all main political actors are in agreement on what needs to be done.

Based on the information uncovered, the principal investigators created six categories of policy issues of interest:

1. (i) Issues of alignment captures spheres of influence (Communist vs. Western), inclusion in broad geographical and cultural groupings (Middle East, Caspian), former colonial empires, membership in alliances, including hosting of bases, and membership in regional blocs (this is a broad category related most centrally to security, but also to ‘civilizational’ choices or groupings).

2. (ii) Issues of armed conflict captures strife, use of force, troop deployment, violence, border clashes, rebel activity, and disputes over territory and nuclear weapons.

3. (iii) Economic issues includes investment, trade, aid, energy, health, reconstruction, and logistics such as roads and ports.

4. (iv) Democracy and human rights issues includes freedoms, repression, elections, persecution of minorities and migration.

5. (v) Left–right issues reflects the use of left–right language.

6. (vi) Global bads captures corruption, transnational crime, terrorism and similar.

We created a dictionary of terms (Appendix D) around those issues and used text analysis techniques to classify their presence in each intervener-targeted-election pair, essentially checking all the highlighted text on which A _gov and B _opp decisions are made. To do so, we built a semantic representation of each issue (that is, a centroid) as the average of the word embeddings of all its seed words (Manning, Raghavan and Schütze Reference Manning, Raghavan and Schütze2008). Then, given the semantic representation of a snippet, we measured its relevance to each issue as the cosine similarity between the two vectors (representing the snippet and the issue).Footnote ¹⁹ The relevance score reflects the extent to which the discussion of the stances of the government and opposition on policies of interest to an outsider matches each of the policy stakes we created.

We can think of the semantic match as the relative weight of the issue in the resulting numeric score. A higher match indicates that a particular issue dominates the discussion, leading to the expectation that this issue is relatively more important or influential in setting the overall tone in relations. For each election and potential intervening power, we rank ordered the different categories according to the degree to which they match, from a ‘most important’ to a ‘least important’ (or least matching) issue. This is equivalent to saying that one issue has the highest congruence, followed by the issue with the next-highest congruence, and concluding with the issue that is least congruent.

Figure 1 shows the distribution of the issues at stake for interveners for each election in our data. Six histograms show the importance of alignment, armed conflict, economic items, democracy and human rights, left–right issues, and global bads such as terrorism and drug trafficking. Each panel overlays two histograms in the interest of space. The results confirm that political alignment and security are dominant themes when candidates discuss relations with a specific external power. Economic matters come in third place. Democracy and human rights issues are of intermediate concern. Left–right divisions and terrorism do not make it to the top of the agenda for most cases.Footnote ²⁰

Figure 1. Policy issue at stake in relations between intervening power and target country: text classification based on nearest-neighbor centroid

Improvements over other Data Sets

The data we provide on political polarization is an advance over the available alternatives. The CMP is a well-known and highly respected effort to assemble a large dataset of what each party said it would do, in written form, before an election was held (Budge et al. Reference Budge2001).Footnote ²¹ Human interpretative coders estimate a variety of position placements for all major political parties based on fifty-six policy categories, which are extracted from the parties’ electoral programs. In addition, scholars have used the manifesto corpus to create alternative policy categorizations and to automatically scale positions based on text. While the dataset is impressive and constitutes an excellent resource for many research agendas, its utility for studying election interventions is limited. For instance, CMP historically has focused on about fifty, mostly developed, countries. It does not cover all the potential issues that may drive a country's relationships with foreign powers. There are some variables related to internationalism, peacefulness and left–right divisions that may be salient for foreign powers. We pick the left–right dimension available from CMP for a comparison to our data as a predictor of interventions. It may be somewhat important in itself, or as a proxy for other dimensions.

UN voting alignment data is also a plausible alternative dataset. UNGA voting patterns have received considerable attention in the literature (Márin-Bosch Reference Márin-Bosch1987; Voeten Reference Voeten2012) and constitute the most widely used measurement of countries’ foreign policy preferences (Bailey, Strezhnev and Voeten Reference Bailey, Strezhnev and Voeten2017). Given the broad range of topics that are voted on every year, and the fact that the large majority of countries are part of the UNGA (Chelotti, Dasandi and Mikhaylov Reference Chelotti, Dasandi and Mikhaylov2019), scholars have been interested in using the expressed preferences from states’ voting behavior – and the resulting voting similarities – as an indicator of state relationships that is comparable across time and space (Kim and Russett Reference Kim and Russett1996; Voeten Reference Voeten2012). While measures for extracting policy preferences and constructing relationship structures are abundant and range from simple (see Voeten Reference Voeten2012 for an overview) to more complex (Magu and Mateos Reference Magu, Mateos and Cherifi2018), recent studies question whether any of these measures can actually capture states’ underlying preferences.Footnote ²² For us, the concern is that issues that are up for a vote at the UN may not correlate with the set of issues at stake in relations between two countries. Furthermore, UN voting is available only for the government, thus ignoring the position of the opposition. Nonetheless, in what follows, we use voting similarity on UNGA votes (in the year before an election) to check whether it predicts US election interventions. We focus on all, rather than only important, votes because these are shown to be highly correlated for most significant purposes.Footnote ²³

Next, we use these datasets to estimate a model of US interventions in the elections of other countries. Appendixes H and I provide a detailed look at how the UN and CMP measurements relate to our Process–Party measurement of polarization for the case of the United States as the intervener.

When does the United States Intervene in Elections?

Historically, the United States has been the most active power to intervene in other countries’ elections, possibly because it is a powerful country and a democracy (and is thus familiar with how elections operate). We use US interventions to illustrate the importance of our theoretical point and empirical measures.

Figure 2 and Table 2 show the linear fit of an ordinary least squares (OLS) model for candidate and process interventions undertaken by the United States against polarization as measured by the Process–Party data.Footnote ²⁴ Polarization is defined as the difference between two –1 to 1 measures, A _gov and B _opp, and thus ranges from –2 to 2. It also shows the same by CMP (left–right split) and against the target government's voting alignment in the UNGA. The UNGA data is the government's voting alignment, normalized and rescaled so that the minimum in the sample is –2 and the maximum is 2, essentially setting the opposition's unobserved alignment to 0. We are forced to do that since no opposition measure is available. As discussed above, we expect an upward slope (β ₁ > 0) for candidate interventions and a downward slope (β ₁ < 0) for process interventions.

Figure 2. US process and candidate interventions: OLS on Process–Party data, UNGA voting alignment, L–R polarization from CMP

Note: pro-democracy interventions are denoted by d+, negative values denote actions that are undermining democracy, candidate interventions are g⁺, pro-government, g⁻, pro-opposition.

Table 2. US process and candidate interventions

Note: OLS on Process–Party data, UNGA voting alignment, L-R polarization from CMP. Process/Candidate Interventions_i = β ₀ + β ₁ ⋅ Divisions_i + ɛ_i. ***p < 0.01, **p < 0.05, *p < 0.1

The Process–Party dataset bears the predictions of the theoretical model, expecting upward-sloping candidate interventions and downward-sloping process interventions over policy polarization.Footnote ²⁵ When the United States wants the government to win (positive and high polarization), it does not promote democracy, and it puts resources behind the governing candidate ticket. When the opposition is preferred, democracy promotion is undertaken to the fullest, and these actions are backed (or accompanied) by attempts to boost the opposition ticket via candidate support. Interestingly, some democracy support is forthcoming even when the United States has no reason to prefer one set of candidates over another (policy polarization of π = 0), suggesting that part of the attraction of this strategy lies in the inherent appeal of promoting democracy abroad. The confidence intervals are narrow, reflecting both precise measures and the larger sample size. Data are available for all elections in the Process–Party dataset. While sometimes the two distinct types of interventions are undertaken separately, in many cases the United States engages in both simultaneously, suggesting that there are complementarities. We could account for this by jointly estimating both outcome variables via seemingly unrelated regressions. We find that standard errors only change by a tiny amount, and that this does not affect overall patterns of significance (cf. Appendix Table F.8).

By contrast, in the CMP and UNGA linear fits, the confidence intervals are wider. In the UNGA model, the slope of candidate interventions is indistinguishable from 0 (suggesting that government alignment does not predict efforts to boost a particular candidate). In the CMP model, the slope of pro-democracy interventions is not distinct from 0, suggesting that left–right divisions are not predictive of an intervening power's commitment to clean rules of the game.

UNGA voting alignment is likely a noisy proxy for the issues at stake in an election. The Process–Party data does not measure solely, or most importantly, foreign policy positions. An outsider may be concerned primarily about domestic issues from the target's perspective, such as drug production, investment policy or human rights abuses. These policies may or may not overlap with the UN's agenda. In addition, voting alignment similarity is based on observed votes: since many items never come up for a vote, the resulting measure may not reflect states’ true conflicts and similarities. India–Pakistan and Greece–Turkey had nearly perfect voting alignment while fighting wars or being on the brink of a war. In addition, the missing opposition position variable in UNGA data may bias the coefficients away from their true slope.Footnote ²⁶

We believe our measurement is an independent source of information that is better suited to studying election interventions. UN alignment may be better for other purposes.

The CMP measurement of polarization does not track the issues over which interventions occur closely enough. This may help explain the wide confidence interval on candidate interventions.Footnote ²⁷ Still, the slope is similar to the one we estimate with the Process–Party data, which may reflect the tendency of the left–right dimension to track anti-Americanism in the West (Chiozza Reference Chiozza2009) and other issues of interest to the United States. In addition, since process-altering interventions may be impractical or impossible in the mostly consolidated democracies covered by CMP, it is not surprising that left–right divisions are not predictive of US decisions to promote more or less clean elections.

Even though we create c and p interventions independently of the polarization variable, we generate an additional measurement of US democratic interventions abroad in order to determine whether our results hold if we use an extraneously generated measurement of interventions. We extract the number of times a country's election is criticized in the Congressional record from 1988 to 2016.Footnote ²⁸ We replicate the democracy-promotion top part of Table 2 with the new measurement in Appendix Table G.11. The Process–Party polarization measurement predicts greater criticism of elections when the opposition has a friendlier stance. Neither the UNGA's alignment nor the CMP's left–right measurements produce similar effects.

Our dataset partly captures covert interventions; we discuss how missed interventions may influence our empirical predictions in Appendix C. We may be underestimating the true effect of divisions on interventions, making this a hard test of the theory.

Logit Models of us Interventions with Process–Party Measures

While OLS provides an easy comparison, our variables may not comply with the types of scaling assumptions needed to run an OLS. For example, policy divisions in Sweden may differ from those in Iran, pro-democracy interventions may require a different threshold than non-democracy ones, and so on. The addition of standard covariates may help us establish the robustness of the relationship.

The main relationship between process and candidate interventions and polarization also holds in an ordered logit model (see Columns 1 and 4 of Table 3), even when we recode process and candidate interventions as dummy variables and use a conventional logit model to check for a relationship (see Columns 2, 3, 5 and 6 of Table 3).

Table 3. Summary table for US interventions

Note: table shows summary statistics for US interventions. A single observation is an election in a certain country that was selected for our random sample of elections.

We also add controls – a dummy variable to indicate whether a target country is developed, a dummy variable measuring whether a target country is a democracy based on the V-Dem polyarchy electoral democracy measurement (Coppedge et al. Reference Coppedge2016), and a post-Cold War dummy variable.Footnote ²⁹ Table 4 shows the summary statistics for our main dataset covering US interventions.

Table 4. Logit models of interventions with Process–Party polarization

Note: US process interventions (d + pro-democracy interventions are indicated by positive values of the outcome variable, d − negative values denote actions that are undermining democracy), and US candidate interventions (positive values indicate pro-government interventions g +). We estimate:

Process/Candidate Interventions_i = β ₀ + β ₁ ⋅ Polarization_i + γ ⋅ Controls_i + ɛ_i.Process/Candidate interventions take values of –1, –0.5, 0, 0.5 and 1. For candidate interventions we estimate an ordered-logit model and binomial-logit models on dummies that take a value of 1 if c > 0 denoted by g + (and accordingly g − for c < 0). For process interventions we repeat this approach and estimate again an ordered logit as well as binomial models using dummies for d + (pro-democracy interventions with p > 0) and d − (for p < 0). ***p < 0.01, **p < 0.05, *p < 0.1

The basic results still hold. As expected, there seem to be fewer interventions in more democratic countries, regardless of the level of polarization. This means that democratic consolidation makes democracy promotion unnecessary, and democracy erosion costly and impractical. Foreign powers are less likely to engage in pro-democracy interventions if they prefer the current government. Higher values of the polarization variable predict a lower probability Pr(d + ), so democracy-enhancing interventions are less frequent. The evidence on democracy erosion Pr(d − ) as a function of polarization is weaker.

Polarization remains a robust predictor of candidate interventions. The polarization variable takes values between –2 and +2, and its standard deviation is approximately 0.9. A logit coefficient for polarization of 2.394 in Column 2 of Table 3 translatesFootnote ³⁰ into an almost tenfold increase in the odds that a candidate intervention will benefit the government if polarization increases by one unit. Similarly, a one-unit increase in polarization reduces the odds that a process intervention will enhance democracyFootnote ³¹ by around 42 per cent.

We also run specifications that control for various regions of the world via dummy variables (see Appendix Table F.9). The coefficient on our ordered-logit models remains similar. There is no region of the world in which candidate interventions are significantly more likely to occur relative to advanced economies; the coefficient for process interventions is positive and significant at the 5 per cent level for Latin America relative to the advanced economies. The dichotomous outcomes Pr(g + ), Pr(g − ), Pr(d + ) and Pr(d − ) display similar coefficients for polarization, but some interesting regularities affect the number of observations and the extent to which these estimations are comparable to the previous estimations. There were no candidate interventions benefiting the opposition ticket in African countries (reducing the number of observations to 159 in the respective estimation). Nor were there any process interventions in advanced economies (reducing the observation count to 166 for both models).Footnote ³²

Additionally, we present evidence based on an alternative variable for candidate divisions extracted from the Process–Party dataset. We create a dummy variable to indicate if the foreign power prefers the opposition and another to denote that it prefers the government. This approach places the fewest assumptions on how much the foreign intervener prefers the opposition over the government when it comes to predicting involvement. Appendix Table F.10 shows the results of these logit models, with dummy variables as both regressors and outcome variables. Our basic intuition (presented above) still holds.

Issues at Stake

Appendix Figure E.4 and Table E.7 report the results on the importance of issues. We expect higher stakes to result in a steeper slope for interventions. We divide the sample into (1) cases in which alignment or armed conflict are the most or second-most important issues and (2) cases in which economic issues, left–right divisions, democracy and global bads assume this ranking. We expect the first group to feature higher stakes (Γ). We do, indeed, find a steeper slope for what we call higher-stakes issues. However, in a fully interactive model, the interaction term between stakes and polarization (not shown) is not significant. Thus we have suggestive, but not statistically significant, evidence that higher policy stakes drive US interventions.

A Look at all Interveners

Figure 3 and Table 5 present the OLS results for all interveners globally in the elections we study, for both the Cold War and post-Cold War periods. The results remain unchanged for the whole sample, with some interesting regularities observed by period (ordered logit yields similar results). Polarization seems to be more strongly driven by policy divisions during the Cold War, possibly due to the higher stakes inherent in superpower competition. Polarization only weakly predicts process interventions during the Cold War; they are much more strongly predicted in the more current period. It is possible that some substitution is at work, with states switching to the more acceptable form of election intervention – affecting democratic processes – after the end of the Cold War. Also, the overall emphasis on democracy promotion (at 0 polarization) is higher after 1990 – a reflection of the growing international concern with clean elections. The latter evidence is consistent with the additional democracy term we introduced in the model. The upward move is statistically significant, suggesting a greater λ in the post-Cold War period. It is also the case, though we do not show it here, that the United States is more committed than any other intervener to promoting cleaner processes.

Figure 3. Process and candidate interventions by all potential (sample) interveners

Table 5. Interventions by all powers

Process/Candidate Interventions_i = β ₀ + β ₁ ⋅ Polarization_i + ɛ_i.

***p < 0.01, **p < 0.05, *p < 0.1

We also use this set-up for two extra robustness checks. Appendix Figure F.6 shows what happens when we replace the human coding of polarization with machine-predicted scores. We see more noise and a wider confidence interval, but the direction and significance are unaltered.

Appendix Figure F.7 evaluates whether elections in which the incumbent is seeking re-election (a presidential election in a presidential system or any parliamentary election in a parliamentary system) have higher stakes for interveners and elicit more activity. There is some evidence to this effect, but it is not very strong. All national-level elections seem to attract foreign interest, possibly because they all matter.

Does it Matter? Pre-election Positions, Post-Election Policies

Does polarization matter? It could be that our measurement is a poor representation of policy positions, or that whoever is elected adapts their position, rendering their pre-election positions irrelevant. The latter may be especially the case for more democratic states, in which incumbents have an incentive to pre-empt opposition challenges through moderation.

We take advantage of two NELDA-provided variables to evaluate this possibility. We combine questions nelda51 (‘Is there a negative change in relations with the US after the election?’) and nelda52 (‘Is there a positive change in relations with the US after the election?’) into a new binary variable coded 1 for cases in which there was any change in relations with the United States after the election.Footnote ³³ We also take the absolute value of polarization as an independent variable. We again divide the sample into democracies based on the V-Dem polyarchy electoral democracy measurement. A score of greater than 0.6 is defined as a democracy (choice of threshold does not matter).

We run a logit model of a post-election policy change in relations with the United States as a function of pre-election polarization in Table 6. We find that growing polarization is associated with a change in relations, regardless of regime type. Absolute polarization is also somewhat lower (0.57) in democracies than non-democracies (0.72). Thus, there is some moderation in democracies but, conditional on polarization, there is a policy change in relations with the United States across regime types. While this discovery calls for a separate explanation, we take these findings as additional validation of the polarization variable we construct and of our general approach.

Table 6. Post-election policies by regime type

***p < 0.01, **p < 0.05, *p < 0.1