Why do parties use primaries?: Political selection versus candidate incentives

See Sect. 1.1 below for a literature review.

Formally, this variable is defined as the share of the electorate that always votes for the party regardless of the quality of its candidate or policies.

The results are similar using the vote share obtained by the party’s candidate in the previous presidential election.

For some examples of work on endogenous political institutions, see Alesina and Trebbi (2004), Ticchi and Vindigni (2010) and Acemoglu et al. (2012). The literature studying the economic effects of political institutions is quite large. For example, Barro (1996), Rodrik and Wacziarg (2005), Papaioannou and Siourounis (2008) and Kudamatsu (2012) study the effect of democracy on growth and well-being, while Persson and Tabellini (2003) and Persson et al. (2003) thoroughly study the effect of electoral rules and form of government on policy choices. In one of the few studies using long-term historical data, Kurrild-Klitgaard (2000) examines the effect of constitutional rules on the incidence of coups in Denmark.

Other possible electoral benefits from primaries may come from its role as informational or commitment devices. For example, Castanheira et al. (2010) discuss a trust effect: parties can use costly primaries to signal the quality of their candidates. Meirowitz (2005) presents a model wherein primaries allow voters an early chance to reveal their preferences. Jackson et al. (2007) argue that primaries allow a party to credibly commit to more centrist policies. In this line, Gerber and Morton (1998) find evidence that U.S. representatives from states with less open primaries take policy positions further from the median voter’s ideal position.

Additional results are presented in an online Appendix available at http://​www.​sfu.​ca/​~faragons/​files/​appendix_​online.​pdf.

Note that the model does not require this effort to generate positive spillovers for the party. The argument could work even if the benefit from this effort is captured only by the politician. The party will, in any case, naturally prefer a higher level of effort than the politician’s optimal because it does not bear the cost of effort.

This excludes actions that affect in the performance internal election, but are not relevant for the general election, as well as actions taken after the candidate secures the party nomination.

The results are similar if we allow for some degree of imperfect observability of types or effort. An extension of the model with imperfect information is available in the online Appendix.

To see this, define G(x)=Pr(max{q ₁,q ₂}<x). Note that G(x)=Pr(q ₁<x)Pr(q ₂<x)=x ². Hence, \(E(\max \lbrace q_{1},q_{2} \rbrace)= \int_{0}^{1} x \, dG=\frac{2}{3}\).

This result is similar to Adams and Merrill (2008) and Serra (2011).

To see this, consider a possible strategy wherein both politicians offer the same level of effort e _a <e _P and have a probability of nomination equal to \(\frac{1}{2}\). Since the probability of nomination increases to one by a small increase in effort e _a +ϵ, a politician will find a unilateral deviation to be profitable, and the strategy profile will not be an equilibrium. A similar argument applies for strategies with different effort levels.

For additional information on Latin American primaries, I refer the reader to Alcántara Sáez (2002), Carey and Polga-Hecimovich (2006: 532–533) and Freidenberg (2003). Martz (1999) and González (1999) provide country analysis for Colombia, Venezuela and Uruguay. For a thorough analysis of the history and organization of the main political parties in some Latin American countries, see Alcántara Sáez and Freidenberg (2002).

The exception is Chile.

One example of this phenomenon has been the decline of the Mexican PRI, after ruling the country for almost 40 years, and the surge of the party PAN. In Venezuela, Peru, Ecuador and Colombia, new parties have displaced old ones and won presidential office, such as Movimiento V República (Hugo Chávez), Cambio 90 (Alberto Fujimori), PSP (Lucio Gutiérrez), Partido Social de Unidad Nacional (Juan Manuel Santos). Similar phenomena can be documented in the rest of the region.

As a rough measure of this heterogeneity, I calculate the standard deviation of the use of primaries between and within parties. The values are 0.268 and 0.277, respectively.

Note that due to data availability, I am unable to distinguish between open and closed primaries in the empirical analysis.

The most important cases are: FREPASO and UCR in Argentina, FSLN in Nicaragua, the Liberal Party in Colombia and, more recently, Concertación in Chile, PRI in Mexico and parties in Uruguay (Freidenberg and Sánchez López 2002).

See Freidenberg and Sánchez López (2002) for additional details on rules and practices to nominate presidential candidates in Latin America.

The online Appendix provides a description of all variables and data sources.

The website addresses are http://​cdp.​binghamton.​edu/​era/​index.​html and http://​pdba.​georgetown.​edu/​Elecdata/​elecdata.​html, respectively.

Note that Latin American countries have a presidential form of government with separate presidential and legislative elections. Both elections, however, may be held at the same time. I prefer to use the seat share in the previous election, instead of the contemporaneous seat share, to reduce concerns of reverse causality between the use of primaries and party electoral performance.

I thank an anonymous referee for pointing out these limitations.

This definition of coalition corresponds to the “narrow” definition used by Kemahlioglu et al. (2009).

See the online Appendix for a group comparison.

The year dummies are included to correct for changes in trends of primary use. The results, however, are robust to the exclusion of these year fixed effects.

The results are robust to using a simpler White correction of the standard errors. I also estimate the relation using discrete dependent variable models. In particular, I use a logit model and conditional logit models with country and party fixed effects. In all cases, the relation remains positive, though it becomes insignificant when using party fixed effects. In this last case, this result may be driven by the drastic reduction of sample size when using the conditional logit model. All these results are available in the online Appendix.

A possible hypothesis motivated by these results is that larger parties may be able to utilize other means of solving internal conflict. In that case, measures of internal division may be less important in determining primary adoption.

In terms of the model, these predictions imply V>0 and \(\frac{dV}{d\lambda}>0\).

Recall that V=V _P −V _NP where V _P and V _NP are the candidate’s vote share when using primary and non-primary, respectively. Formally, the estimate of V is ϕ ₀+ϕ ₁ SEATSHARE1.

These results echo the findings of Carey and Polga-Hecimovich (2006). Using a similar dataset but a different identification strategy, they find that primary-nominated candidates obtain a vote share between four and six percentage points higher than candidates selected in other ways.

There are, however, other possible reasons why the primary bonus may increase with partisan support. For example, the gains from a primary may be a relative increase in partisan support instead of an absolute gain in vote share. More broadly, the results would be similar if the marginal benefit from using primaries increases with partisan support. While I cannot rule out this possibility, I explore the robustness of the results to the inclusion of the interaction of primary with a proxy for party size at the moment of the election, i.e., after the nominating decision was announced. As such a proxy, I use the party’s seat share obtained in the legislative elections contemporaneous with the presidential election (SEATSHARE). The results, available in the online Appendix, are similar.

A similar graph for the increase in the probability of winning the presidential election is available in the online Appendix.

To construct the confidence interval, I first calculate the standard errors of each point estimate using the estimated covariance matrix. Then, I add and subtract 1.64 S.E. to the point estimates.

See Sect. 3.3 for a more detailed discussion.

I refer the reader to Kemahlioglu et al. (2009) for a discussion of these primary determinants.