Assumptions

A way to deal with these commitment problems is to place the machine–voter interaction in a dynamic context. To model the interaction between machine operatives and voters as a repeated game, we have to make certain assumptions. First, we have to assume that parties can monitor individual voters' actions and condition rewards on their inferred votes. Second, we have to assume that both sides perceive the interaction as ongoing indefinitely into the future.

The assumption that machines can hold voters accountable, that they can monitor individuals' votes (even if imperfectly) and make rewards contingent on the voter's support, departs from the implicit assumption of redistributive theorists. They assume that a member of a favored group will receive private rewards whether or not he votes for the party; individual voters are anonymous and therefore free from the party's retribution should they defect. The premise that voting is a private and anonymous act may have discouraged formal theorists from modeling these interactions as repeated games; repeated games generally rely on each player being able to observe the actions of the other in the previous round. The assumption that voting is anonymous is appropriate in most advanced democracies, but not necessarily in the historical context of political machines or in many new democracies today.⁶

There are two kinds of private information about the voter that are useful to the party: his actions—which party he votes for—and his type—his partisan predisposition in relation to the two parties.⁷

Certain voting technologies allow parties to monitor individuals' votes. The recent historiography of U.S. machines deepens our appreciation of these technologies. Until the introduction of the Australian ballot in the United States, in most states in 1891, parties produced “ticket” or “coupon” ballots, ones that listed only their candidates. To monitor which party's ballot the voter was using, parties printed ballots on paper of different weights or colors. Voters deposited the ballot directly in the ballot box, under the watchful eye of party operatives, without first concealing it in an envelope (for descriptions, see Keyssar 2000; Reynolds 1988). Reynolds (1980, 193) reports that New Jersey's early automatic voting machines, introduced in 1890, made clicking noises that allowed party officials standing nearby to detect the voter's selection. And operatives from the Philadelphia Republican party in the late nineteenth and early twentieth century offered to fill out ballots on voters' behalf (McCaffery 1993).

Voting practices and technologies undermine the anonymity of the vote in contemporary developing democracies, as they did in U.S. machine cities, even where the Australian ballot is in use and where voting is, in a narrow sense, secret. In her description of contemporary India as a “patronage democracy,” Chandra (2004) notes that parties designate polling agents to observe the progress of voting. Polling agents are “usually men from the village itself, or from close by, who know the identity of each voter. While they do not witness the actual vote, they know who shows up to vote and can report on turnout figures” (139). Chandra reports that Indian parties could undermine voters' anonymity by emptying boxes and counting the returns at frequent intervals over the course of an election day. (An electoral reform in 1994 outlawed the practice.) To cite another example, in the 2003 Russian Duma elections, international observers reported “significant problems relating to the secrecy of the vote, with open voting in 30% … of polling stations … polling officials and party observers were seen to be actively encouraging persons to vote outside of polling booths” (Organization for Security and Cooperation in Europe 2003).

Certain party–organizational structures allow parties to discern individual voters' types—their predisposition for or against the machine. The typical political machine (or clientelist party) is bottom-heavy, decentralized, and relies on an army of grassroots militants. Voters in today's democracies in the developing world are frequently geographically immobile, living in neighborhoods where they grew up and where family members and close acquaintances live. Some of these familiar neighbors work as operatives for political parties. They therefore know much about an individual that shapes his partisan attachments: his job, associational membership, parents' ideological inclinations, and public statements about parties and policies. It is also hard for voters to dissemble before people they've known all their lives: as one grassroots party organizer in Argentina explained, you know if a neighbor voted against your party if he can't look you in the eye on election day.

Information about individual voters' partisan predispositions helps the machine make inferences about how individuals vote and whether they are good candidates for vote buying. For instance, the model in the next section shows that voters who are predisposed in favor of the machine on partisan or programmatic grounds cannot credibly threaten to punish their favored party if it withholds rewards. Therefore the party should not waste rewards on them. The model also shows that voters who are strongly opposed to the machine will not trade their votes for rewards. A machine can compensate, to some degree, for an effective secret ballot if it can distinguish strong opponents from people who oppose it more moderately, or strong loyalists from people who are indifferent about whom to vote for.

Argentina, the country from which I present evidence, combines a balloting system that gives parties greater control over voters than does the Australian ballot, a social structure of reduced anonymity, especially among the poor, and party organizations that help parties monitor voters. These features contribute to a widespread perception among Argentine voters and party operatives that voting is a less than fully anonymous act. As one grassroots party organizer explained, “Anyone who's militating in the streets, you know who's with you and who's not with you.”⁸

Voting technologies in Argentina also reduce the anonymity of the vote. Argentina has the secret but not the Australian ballot.¹¹

The practice of handing out ballots basically serves as a method of monitoring and influencing how people vote. One Peronist organizer explained in an interview how the party used the ballots. “The most important thing is to go look for people and give them the ballot. You give them the ballot in the taxi [which the party has hired to transport them to the polls]. Then no one has time to change their ballots for them [i.e., give them a different ballot. After taking voters into the polling place] you put them on line to vote … Then they don't have a chance to change the ballot. Only if they're really sneaky and they change it inside the voting booth.”¹²

In sum, my first assumption is that machines can effectively, if imperfectly, monitor the actions of their constituents.

A second assumption needed to model machine politics as a repeated game is that all players foresee the game continuing into the future. It is entirely appropriate to think of the interactions between machine operatives and their constituents as repeated over many iterations; the more artificial assumption would be that these are one-shot or short-lived interactions. Machines and clientelist parties are effective to the extent that they insert themselves into the social networks of constituents. The grassroots party operative is a long-time neighbor of the people she tries to mobilize. In Latin America, clientelist parties of renown have been long-standing organizations, deeply enmeshed in working-class communities: Peru's Partido Aprista Peruano (APRA), founded in the 1920s, Mexico's Institutional Revolutionary Party (PRI), founded in the 1930s, Argentina's Peronists, founded in the in the 1940s.

The repeated-play assumption may be most appropriate in countries where parties are old, even if the democracies in which they compete are new. The three democracies just mentioned are new: Peru and Argentina redemocratized in 1980 and 1983, respectively, and Mexico democratized for the first time in 2000. Yet clientelist parties in all of them are old. The repeated-play assumption may be less appropriate in new democracies where the major political parties are also young and hence less enmeshed in social networks.

When parties that are not enmeshed in social networks try to buy votes with private inducements, voters greet their efforts with skepticism. In connection with research I conducted in Lima, Peru, I observed the reactions of people in a working-class neighborhood to a soup kitchen that a political party established in 1985, shortly before national elections (see Stokes, 1995). Soup kitchens were familiar in the neighborhood: Catholic activists and women's organizations ran some and the local mayor's office supported them. But when residents saw an outsider party set up a soup kitchen they predicted that it would disappear after election day. They were unmoved by the sponsor-party's implicit appeal for electoral support. And they were right: the soup kitchen did disappear right after the election.

In the Argentine case, furthermore, it is appropriate to assume that parties and voters see their interaction as extending into the indefinite future; even if they could imagine hypothetical circumstances in which it might end (in the event, e.g., of a military coup), at the time of any given election since the return to democracy in that country, few would have anticipated a particular moment when it would end. The perception of an interaction with no identifiable stopping point makes it reasonable to model this as an infinitely repeated game.

To capture the repeated-play dynamic of machine politics, it is necessary to depart in a third way from received models of redistributive politics. These models assume that the machine's ability to reward voters for their support depends on its winning elections.¹³

To summarize, my key assumptions are that machines can monitor voters' actions and that both sides foresee their interaction extending indefinitely into the future. The latter assumption implies that machines don't lose their ability to distribute goods when they find themselves in opposition.

The Model

I begin with a one-shot game in which a person's vote is assumed to be perfectly observable by political parties. Let the ideological position of the machine in a one-dimensional policy space be represented by x₁, the ideological position of the opposition by x₂, and x₁<x₂. Let x*≡(x₁+x₂)/2 be the midpoint between the two parties (see Figure 1). Let the voters' preferences be given by

where v_i={x₁, x₂} represents a vote for either the machine or the opposition, x_i represents voter i's position on the ideological spectrum, and b_i={0, b} represents the value to the voter of the reward offered by the machine in exchange for votes, relative to the value of voting according to the voter's preferences. Thus −(1/2)(v_i−x_i)² represents the expressive value of voting for one of the two parties. If the machine does not offer a gift, then b_i=0 and the voter votes for the machine if −(x_i−x₁)²≥−(x_i−x₂)², or if x_i≥x*. That is, if there is no gift the voter supports the party that falls closest to the voter on the ideological or programmatic dimension. If the machine offers a gift of b>0, the voter will vote for it if

or

or

The normal form of the stage game is depicted in Table 1. In the Table, the machine is represented as expending b when it pays a reward, and gaining v when it receives a vote.

The Location on a Spatial Dimension of a Political Machine(x1), Its Opponent (x2), the Median Voter (x*), and a Hypothetical Voter (xi)

Define voters for whom x<x* as Loyal voters (see Figure 2). Loyal voters' dominant strategy is to vote for the machine. Define voters for whom x>x+b/(x₂−x₁) as Opposition voters. Opposition voters will oppose the machine even if offered b to change their votes. Define voters for whom x*<x<x*+b/(x₂−x₁) as Weakly opposed voters. Weakly opposed voters prefer to vote against the machine in the absence of a reward, but prefer to vote for the machine if doing so brings them a reward. If the value of the vote to the machine exceeds b, the machine and the Weakly opposed voter are in a prisoners' dilemma. Table 2 gives the game between a Weakly opposed voter and a machine, with simplified payoffs that make clear the prisoners'–dilemma structure of the game.

Types of Voters by Their Location on a One-Dimensional Policy Space

Next, I assume an infinite sequence of elections and model the interaction between the machine and a Weakly opposed voter as an iterated prisoners' dilemma with one-sided uncertainty.¹⁴

Returning to the model, if the voter votes against the machine, I now assume, the machine observes the negative vote with a probability p. Voters discount the future by a discount factor β, which falls on the interval [0, 1]. The condition for a subgame-perfect equilibrium (SPE) in which the Weakly opposed voter receives the reward and votes for the machine, supported by a grim trigger strategy should the voter be observed to renege, is

In other words, to sustain cooperation, the value to the voter in the current and all subsequent periods of voting for the machine and receiving a reward must equal or exceed the sum of the payoff from defecting in the current period plus (1) avoiding detection and returning to cooperation in the next and subsequent periods (with probability p), or (2) being caught and, in all subsequent periods, voting against the machine but foregoing rewards (with probability 1−p).

Inequality [1] simplifies to

where

Hence, the set of voters who would sell their votes in exchange for a private benefit is the set whose ideal point, x_i, satisfies

Lambda falls on the [0, 1] interval. Lambda is an increasing function of the discount rate (β) and of the probability of a defector being caught (p). If p=0 (there is no possibility that the machine would observe a defection by the voter), or if β=0 (the voter cares nothing about future consumption), then inequality [2] reduces to x_i=x*. In these cases the machine can buy the votes only of voters who are indifferent, on ideological grounds, between the parties.

Loyal voters do not meet the condition in [2]. As illustrated in Figure 2, for Loyal voters x_L<x*. Intuitively, Loyal voters who want to extract private rewards from their preferred party would, under the grim trigger, have to threaten to vote against the party forever if the machine denied them a reward once. Such a threat would lack credibility: the party knows that the Loyal voter, even without rewards, is better off cooperating forever than defecting forever.¹⁶

Weakly opposed voters (and indifferent voters, where x_i=x*) are the only types whose policy ideal points make them potential vote sellers.¹⁷

Inequality [2] implies four comparative statics:

• As the ideological distance between the two parties (x₂−x₁) shrinks, the potential for vote buying grows. Intuitively, when the two parties are ideologically or programmatically close, there is less at stake for the voter in the decision of which to vote for, and the value of the private reward becomes more salient.
• As the value of the private reward (b) relative to the value of voting in accordance to one's policy or ideological preference increases, the potential for vote buying increases. The reward must be worth a lot to the voter. But its value to the machine must be less than the value of a single vote—not very much. This suggests that, given decreasing marginal utility from income, machines will target poor voters.
• The more accurately the machine can monitor voters, the greater the potential for vote buying (λ is an increasing function of p). This accuracy is a function of the technology for monitoring voters' actions and of the machine's organizational structure.
• Among its core constituents—those whom it can observe well—the machine is most effective when it targets Weakly opposed voters (for whom x*≤x_i≤x*+λ(b/x₂−x₁)), rather than Loyal (x_i<x*) or Opposition voters (x_i>x*+b/(x₂−x₁)) voters.

Poverty and Vote Buying

I discuss five pieces of evidence from the survey that lend support to my theory of machine politics. The first has to do with the effect of poverty on a voter's willingness to sell his or her vote. The formal model analyzed earlier predicts that vote buying is more easily sustained, all else equal, when the voter values the private reward relatively highly but the party values it relatively little. The picture this paints is of parties giving minor benefits to voters who are poor enough to value them highly—a picture consistent with much of the qualitative literature on machine and clientelist parties. To cite just one of many examples, Wilson and Banfield (1963) explain that U.S. machines operated in a city's “river wards,” where working-class residents lived, but not in the “newspaper wards,” where middle-class residents lived.

Table 3 reports regression estimates of the likelihood of a clientelistic response to the set of questions discussed earlier, including whether the respondent received a private reward from a party. The negative and significant coefficients on Income, Education, and Housing quality variables show that poverty predicts clientelism. To illustrate the effect, the simulated expected probability that a wealthy person (one with the highest income, education, and housing-quality level) would have received a reward and acknowledged that it influenced her vote is 0.2%. The probability that a poor person (one with the lowest income, education, and housing-quality level) would have received a reward and allowed his or her vote to be influenced by it is 65 times greater: 13%.²¹

In sum, political machines buy the votes of poor people in Argentina.

Monitoring Voters

Types of Voters and Vote Buying

A fifth piece of evidence speaks to the question, What types of voters do machines pursue? My theoretical prediction was that machines focus their vote-buying efforts on people in the middle of the distribution of partisan predispositions: ones who are indifferent about whether to vote for or against the machine (x_M=x*), and ones with a weak predisposition against it (x*≤x≤x*+b/(x₂−x₁)). Machines will avoid voters who are loyalists or strong opponents.

We asked respondents their opinions of the Peronist party, Argentina's preeminent political machine. We asked them to choose among “very good,” “good,” “bad,” and “very bad” as their answers. Figure 3 displays the percentages of people who received or did not receive handouts by opinions of the Peronist party.

Opinions of Peronists Among Recipients and Nonrecipients of Rewards

A striking finding, and one that conforms to the theoretical prediction, is the small proportion of those who rated the party “very good” who received rewards. Three times as many people who did not receive rewards as those who did receive them rated the Peronists “very good” (31% vs. 10%). The Peronist party turned away from its strongest loyalists when it gave out private rewards. (The difference is all the more striking given that one might anticipate some endogeneity of perceptions of the party: people who receive rewards from it might be more prone, because of the gift, to rate the party “very good.”) Another aspect of the findings that accords with my model's predictions is that many more people who rated the party “bad” received rewards than those who rated it “very bad.”

In some ways, however, the findings do not accord with the predictions. Recipients of rewards were concentrated in the “Good” category: nearly 60% of those who received handouts from the Peronists saw it as a good party. These findings are inconsistent with the theory if we think of people who called the machine “good” as falling somewhat to the left of the median ideal point (x*) in Figure 2 and hence as being weakly predisposed in the machine's favor. Recall that, in theory, even voters just mildly predisposed in the machine's favor would not be able to credibly threaten to punish the machine if it defected and therefore would not, in repeated play, be able to induce the machine to pay them rewards. And we might have expected a relatively larger proportion of voters who rated the Peronists as “bad,” and hence who were weakly opposed to it, to receive rewards. Similarly, note that the regression models in Table 3 show that, controlling for other factors, Peronist sympathizers were significantly more likely to receive rewards.

One explanation for the slippage between the theory and the evidence is that our survey did not offer people the option of indicating true indifference about the Peronists. Some people who chose the “good” option might in fact be closer to indifferent. And some people who were close to indifferent, prerewards, might have called the party “bad” but, because of the reward, been nudged into seeing it as “good.”

The finding may also suggest a dynamic that goes beyond the model. Political machines organize by neighborhood and district, and they do more than just give out tactical rewards. They also proselytize. Although their proselytizing, in a competitive setting such as Argentina's, is not perfectly successful, to the extent that it is successful at all we expect the distribution of voter types in areas of machine organizational penetration to be skewed toward machine supporters, weak and strong. In other words, we expect organizational penetration by the party to increase not only the efficiency with which it distributes rewards and its ability to monitor voters, but also its partisan support (as Cox and McCubbins 1986 assume). If organizational penetration increases partisan support, then the machine will target its supporters more than its opponents simply because it has greater access to them. Whatever the explanation for this anomaly, the evidence from Argentina does show unambiguously that, among core constituents, the machine discriminates against its most ardent supporters.