The Party System

With nearly 40 per cent of Argentina’s population, the Province of Buenos Aires (PBA) is the largest unit in Argentina’s federation and a central battleground in national politics. Its 135 municipalities vary significantly in size and influence: in 2011, the population of La Matanza (890,000 registered voters) and Lomas de Zamora (451,000) exceeded that of several Argentine provinces, while the smallest municipality had just 1,701 registered voters.Footnote ⁴

Argentina’s party system has long been defined by a persistent divide between Peronists and non-Peronists – two broad, heterogeneous, and highly factionalized camps that have shaped the country’s politics since the 1940s. But while the two major political vehicles for these camps – the Peronist Partido Justicialista (PJ) and the non-Peronist Unión Cívica Radical (UCR) – have historically dominated, voter backlashes and internal crises have periodically created openings for third-party competition, and rival factions have sometimes opted to compete outside the main party structures.

After regaining control of the province in 1987, the PJ established one of the largest clientelistic networks in the country (Levitsky Reference Levitsky2001; Palermo and Novaro Reference Palermo and Novaro1996), which persisted despite significant transformation in the provincial and national party systems (Calvo and Escolar Reference Calvo and Escolar2005; Leiras Reference Leiras2007). The 2001 economic crisis was especially dramatic for the non-Peronist space (Lupu Reference Lupu2016; Torre Reference Torre2003), as some of the most popular UCR politicians left the party to create their own forces, while new challengers emerged on the center-right. Controlling the presidency allowed the PJ to remain relatively united in Buenos Aires (Cherny et al. Reference Cherny, Feierherd and Novaro2010), while the UCR formed a coalition with Propuesta Republicana (PRO), a center-right party founded after 2001, to beat the PJ in 2015, 2017, and 2021. The rise of Javier Milei’s La Libertad Avanza (LLA) in 2023 introduced a new political force, but this party has struggled to establish a solid organizational structure in Buenos Aires and remains weak at the municipal level.

Yet despite losing the national presidency in 1999, 2015, and 2023, the PJ only relinquished control of the provincial governorship during 2015–19. Its dominance is especially marked in the Conurbano, the densely populated industrial belt surrounding the City of Buenos Aires that is home to ≈ 75 per cent of the province’s population. Nonetheless, competitive races in the Conurbano are not uncommon, especially in years when the non-Peronist camp has an attractive presidential candidate. And the UCR remains a strong competitor in the province’s Interior – a region comprising a rural hinterland and mid-sized cities – where elections are highly contested between both major parties.

In sum, municipal elections in Buenos Aires are characterized by intense electoral competition, driven by both the presence of third-party forces and episodic splits within the two major parties. While the PJ and UCR (plus allies) typically secure around 80 per cent of the vote (see the bottom left panel of Figure A2 in the Appendix) and the margin between the first- and second-placed party in the general election ranges between 15 percentage points in midterms to 18 percentage points in concurrent elections (see Table A1(b)), the largest party often falls short of an outright majority, many races are decided by narrow margins – 18.7 per cent by less than 5 percentage points and 35.2 per cent by less than 10 percentage points – and the effective number of parties measured by the Golosov index (Golosov Reference Golosov2010) ranges between 2.2 and 2.8 on average (Figure A2). Third parties rarely win mayoral elections – just 8 per cent, compared to the PJ’s 56.1 per cent and the UCR’s 35.9 per cent – but they can tip the balance in favor of (or against) one of the two major parties, making coordination essential.

Electoral Rules

Municipalities are governed by a mayor and between 6 and 24 councilors. All municipal authorities serve four-year periods and first faced term limits in 2025. Local councils are renewed by halves every two years: in concurrent years (2011, 2015, 2019, and 2023), both the mayor and half of the council are elected simultaneously; two years later, the other half of the council is elected in a midterm election (2013, 2017, and 2021). Mayors are elected by plurality rule, whereas council seats are allocated using the largest remainders method with a Hare quota. The combination of small districts (see Figure A1) with a high threshold (one Hare quota) means that the two or three largest parties often capture most of the seats. In addition, in concurrent elections voters are forced to support a mayor and a list of councilors from the same party, further advantaging large parties.

Voting is mandatory; sanctions are rarely enforced, but turnout is generally upwards of 75 per cent (see Figure 1). Between 2005 and 2023, municipal elections always took place on the same day as national and provincial races. Thus, while mayors are well-known and important political players, municipal elections are often shadowed by presidential and gubernatorial contests. This logic is strengthened by an electoral technology that discourages split-ticket voting (Barnes et al. Reference Barnes, Tchintian and Alles2017): parties print their own ballots, and often distribute very long sheets of paper listing the party’s candidates for all offices. While voters may physically cut these in order to vote for different parties for different offices, many simply vote for all the candidates aligned with the presidential (or gubernatorial) candidate of their choice. The point is that voters probably pay more attention to national and provincial elections than to municipal ones, and therefore our estimates should be interpreted as a lower bound.

Figure 1.

Evolution of the main variables over time: margins between the most voted parties on top; outcomes in the middle and below. The wider vertical lines indicate concurrent elections.

Between 2011 and 2023, the EPAOS significantly altered electoral dynamics at the national, provincial, and municipal levels (Vallejo Reference Vallejo2025). These were promoted by the national government shortly after narrowly losing the 2009 midterms, with the twin goals of democratizing intraparty life by weakening the nominating power of party elites and improving representation through reduced electoral fragmentation (Abal Medina et al. Reference Abal Medina, Tullio, Eberhardt, Mutti and Torres2023).Footnote ⁵ Shortly afterwards, many provinces, including Buenos Aires, adopted a very similar system for provincial and municipal elections.Footnote ⁶ Citing the need to reduce public spending, the government of Javier Milei negotiated with the opposition to waive the primaries for the 2025 election; this means that the law remains in the books and in principle the primaries will be employed again in 2027. The province of Buenos Aires quickly followed suit.

The primaries take place between 9 and 11 weeks before the general election. Only parties whose combined vote share surpasses 1.5 per cent of positive votes are entitled to contest the general election. Voting is mandatory, with voters restricted to selecting a single party and a single list, including the sole official list if the voter’s preferred party features no internal competition.Footnote ⁷ Only the most popular (or the only) list of each party may advance to the general election. Intraparty competition is thus allowed but not mandated: just 23.3 per cent of parties in our sample featured a competitive primary, and in half of those cases the most voted faction won by an intraparty margin of at least 25 percentage points.Footnote ⁸ During the period of analysis, provincial and national primaries always took place on the same day, following similar rules (though see note 6).

Roughly a third of parties fail to pass the 1.5 per cent barrier, but their combined vote share falls below 2.5 per cent of positive votes (see Figure 2). Voluntarily dropping out is rare: just 1.8 per cent of the 4,421 parties that surpassed the threshold withdrew from the race.Footnote ⁹ The EPAOS are thus quite different from run-off systems, in which the election may be decided in the first round or in a second round that is restricted to the top two (sometimes the top three or top four) vote-getters. The point is that EPAOS work more as a comprehensive and easily available pre-election poll than a mechanism for filtering out parties: the combination of mandatory participation with a structured format serves to forecast the frontrunner, the most viable challenger, and the distribution of electoral support among parties more generally. Parties generally retain their primary-election ranking in the general election or move at most just one position up or down (see Figure A3). But this also means that primary results are not fate, and voters and party elites have room for adjusting their behavior and strategies.

Figure 2.

Proportion of parties participating in the primary that contested the general election. ‘Qualified’ means that a party obtained at least 1.5 per cent of positive votes in the primary.

Graphical Analysis

Figure 1 shows the evolution over time of both the two margins of interest (1 v. 2 and 2 v. 3) as well as the outcome variables: turnout; the proportion of positive votes; the combined vote share of the two largest parties; and the Golosov (Reference Golosov2010) index.Footnote ¹² On average, the leading party surpasses the trailing one by 17 percentage points, with substantial variation between municipalities. The difference is somewhat larger in the primary. In contrast, the 2 v. 3 difference increases in the general election, especially during 2001–17, when there was more uncertainty about the identity of the second-placed party. Both turnout and the proportion of positive votes are noticeably lower in the primary.

The combined vote share of the two largest parties and the Golosov index indicate that electoral fragmentation is larger in the primary, especially between 2011 and 2017. This is consistent with voters using primary results to identify, and vote for, the two frontrunners.

Figure 3 examines how the margin between the leading and trailing parties in the primary affects the change in outcome variables between the primary and the general. A closely fought primary increases both turnout and the share of positive votes, though the first relationship is limited to concurrent elections. A smaller margin between the two most voted parties also increases support for the two largest parties: in concurrent years, their combined vote share increases by 4.2 percentage points if they received the same number of votes, but this decreases by 0.13 percentage points for every percentage point difference between them (see Figure 3(c)). In midterm years, the effect is smaller (3.4 and minus 0.093 percentage points, respectively), but still substantial. In very close elections there are between 0.33 and 0.38 fewer effective parties, but this number increases by 0.012–0.013 for every percentage-point difference between the frontrunner and the runner-up.

Figure 3.

Primary closeness and change in outcome values between the primary and the general.

Regression Analysis

By looking at the change between the primary and the general election, the plots in Figure 3 account for the fact that outcome values in a given municipality-year may be abnormally high (or low) for reasons that have already manifested in the primary. An alternative strategy is to include municipality and year fixed effects, and thus we estimate models of the form

(1) $$y_{m,t}^{\rm{G}} = \beta \cdot {\rm{margin}}_{m,t}^{\rm{P}} + {\mu _m} + {\delta _t} + {\varepsilon _{m,t}},$$

where $$y_{m,t}^{\rm{G}}$$ is the outcome (in levels) measured in the general election in municipality m in election year t, $${\rm{margin}}_{m,t}^{\rm{P}}$$ is the percentage point difference between the leading and trailing parties in the primary, and μ _m and δ _t are municipality and election year fixed effects, respectively. Since the set of parties participating in the primary and the general election may differ, when computing vote percentages and victory margins in the primary we only include the vote totals of the parties that qualified to take part in the general election (that is, that surpassed the 1.5 per cent threshold) in the denominator.Footnote ¹³ To account for floor and ceiling effects, in some specifications we control for the outcome value in the primary, $$y_{m,t}^{\rm{P}}$$ . We cluster standard errors by municipality. All data come from the province’s electoral authority.Footnote ¹⁴

Table 1(a) shows the overall effect for the entire sample. For every percentage point increase in the margin between the leading and trailing parties in the primary, turnout in the general election goes down by 0.021 percentage points – 0.022 if controlling for the lagged outcome – both statistically significant estimates. These effects may seem comparatively small (cf. Arnold Reference Arnold2018; De Paola and Scoppa Reference De Paola and Scoppa2014; Fauvelle-Aymar and François Reference Fauvelle-Aymar and François2006; Garmann Reference Garmann2014; Indridason Reference Indridason2008; Simonovits Reference Simonovits2012), but this is likely due to the baseline turnout already being high.Footnote ¹⁵ The effect on positive votes in columns (3) and (4) is also negative and significant, but between two and three times larger in size. This is confirmed by the standardized estimates reported in Appendix Table A3: the effect of a (within-municipality)Footnote ¹⁶ standard deviation increase in the margin of victory in the primary is two to three times larger for positive votes than for turnout.

Table 1.

Between-party closeness in the primary and general election outcomes

Note: Ordinary least squares (OLS) regression estimates. Each panel–column combination reports a different specification. The outcome is always measured in the general election. margin^P is the difference between the percentage of votes of the leading and trailing parties in the primary election, including only parties that classified to the general election in the denominator. Panel (b) reports separate marginal effects for concurrent and midterm elections; the ‘p-value of the difference’ indicates whether these are statistically different from each other. Standard errors clustered by municipality in parentheses.

Where do these additional votes go? Columns (5) and (6) show that for every percentage point increase in the margin between the leading and trailing parties, the combined vote percentage of the two largest parties goes down by between 0.07 and 0.12 percentage points, depending on whether the outcome in the primary is included as a control. Surprisingly, column (7) indicates that this results in a smaller Golosov index – that is, lower electoral concentration in the general election as the primary becomes less competitive – but column (8) indicates that when accounting for the level of concentration in the primary, the sign switches and the effect becomes positive as expected. The standardized results in Appendix Table A3(a) lead to similar conclusions.

Panel (b) compares concurrent v. midterm elections. To make the results more intuitive, we report the marginal effects of the margin of victory in concurrent and midterm elections, as well as the p-values for the difference between the two. The results are much larger – and more likely to be significant – in concurrent elections. That said, the difference in the marginal effect between concurrent and midterm years is only statistically significant at conventional levels for the combined vote share of the largest parties and the Golosov index. Using council size as a proxy for municipality size, Table A2(d) in the Appendix further shows that voters in small districts are more responsive to electoral closeness: in municipalities with six councilors, the impact of closeness on turnout is two to three times larger than in the overall sample. This effect diminishes almost monotonically with increasing council size and becomes negligible in municipalities with 16 or more councilors. The relationship between council size and positive votes is less predictable, however, and there is no clear pattern between council size and the magnitude of the estimated effect for the other two outcomes. The standardized estimates in Appendix Table A3 also support this interpretation.

Robustness

Table A4 in the Appendix shows that for all the variables of interest, the value observed in the primary is a much better predictor than the values from the general elections that took place two or four years before. Looking at the change in the outcome variable between the primary and the general election (Table A5), measuring vote percentages in the primary without removing parties that did not pass the 1.5 per cent threshold from the denominator (Table A6), or taking the natural logarithm of raw votes or the Golosov index instead of vote percentages (Table A7) does not change our findings either. Accounting for intraparty competition in the primary by either (a) splitting the sample depending on which of the top two parties had multiple lists (Table A8) or (b) defining the explanatory variable as the margin between the biggest factions within each of the top two parties (Table A9) also produces similar findings. Neither the incumbent party’s distance to a majority in the council (Table A10) nor district magnitude (Table A11) change the result for midterm elections.Footnote ¹⁷

Identification

Determining if a party does better in the general election solely by virtue of having finished in a higher-ranked position in the primary is problematic insofar as better-ranked parties are more popular, nominate more attractive candidates, or control more resources. We thus employ a regression discontinuity (RD) design, comparing parties who finished first instead of second (or second instead of third) by a small margin. Following Calonico et al. (Reference Calonico, Cattaneo and Titiunik2014), we estimate this effect non-parametrically, fitting a separate regression at each side of the cutoff point of zero and weighting observations close to the cutoff more heavily. For a given outcome variable, we choose the bandwidth that minimizes the estimates’ asymptotic mean squared error (MSE). Since we include two observations for every election, both the density of the running variable and all election-specific characteristics are perfectly balanced by design. We cluster the standard errors by election year to account for the dependency across observations.

Graphical Evidence

The regression discontinuity plots in Figure 4 show the relationship between a party’s margin of victory in the primary and its probability of winning or its vote percentage in the general election. The plots on the left show that the larger the first-placed party’s margin, the more likely it is to win the election and the higher its expected vote share, but there is no visible ‘jump’ at the discontinuity: finishing first in the primary does not confer an electoral advantage of its own in the general election. In contrast, the plot in the top right corner shows that there is an advantage of finishing second: the third-placed party rarely wins the election, but the second-placed one emerges as winner between 10 per cent and 20 per cent of the time, and the difference begins to show up right at the discontinuity. There is no visible effect for vote shares, however. Figures A9 and A10 in the Appendix suggest that these results are driven by concurrent elections and small municipalities.

Figure 4.

Mimicking variance regression discontinuity plots with quantile-spaced bins (Calonico et al. Reference Calonico, Cattaneo and Titiunik2015) showing the relationship between the margin in the primary and the probability of winning (top) or the expected vote share (bottom) in the general election.

RD Results

Table 2(a) presents the results for the full sample. Finishing first in the primary has a negative and sizable – minus 9 percentage points – effect on the probability of winning the general election. The estimate is not statistically significant, probably due to low statistical power: as the last three columns of the table show, we generally have 80 per cent power to find an effect as large as a standard deviation of the outcome in the control group (s.d. _C ) and sometimes one half as large (s.d. _C /2), but our RD estimates (|τ̂ _RD |) are usually much smaller than that.

Table 2.

Regression discontinuity (RD) estimates: effect of primary ranking on general election outcomes

Note: Sharp (conventional) RD estimates, with robust confidence intervals (CIs) and p-values based on the MSE-optimal bandwidth proposed by Calonico et al. (Reference Calonico, Cattaneo and Titiunik2014), using a triangular kernel and clustering the standard errors by election year. The running variable is the primary election margin between the first- and second-placed parties (odd-numbered rows) or the second- and third-placed ones (even-numbered rows). Only parties that classified to the general election are included in the denominator. The last three columns report how much statistical power the model has to detect an effect that is as large as (a) a standard deviation of the outcome variable in the control group (s.d. _C ); (b) half as much; or (c) equal in absolute value to the one we actually estimated (|τ̂ _RD |).

In any case, the first-place advantage documented in municipal elections in Brazil (Lucardi et al. Reference Lucardi, Micozzi and Vallejo2023) or in legislative elections in France and other European countries (Granzier et al. Reference Granzier, Pons and Tricaud2023) does not extend to Buenos Aires. But in line with Anagol and Fujiwara’s (Reference Anagol and Fujiwara2016) findings for Brazil, Canada, and India, finishing second instead of third provides a 9 percentage point increase in the probability of winning the general election (p=0.04). The estimates for vote shares go in the expected direction – a 1.1 percentage point decrease and increase, respectively – though the small effect sizes and insufficient power means that neither effect is significant. Finding stronger regression discontinuity estimates for winning probabilities than for vote shares is common (see Granzier et al. Reference Granzier, Pons and Tricaud2023; Lucardi et al. Reference Lucardi, Micozzi and Vallejo2023). On the one hand, we are comparing higher- and lower-ranked parties with very similar vote percentages, and thus even a small change in vote shares may translate into an appreciable increase in winning probabilities. Alternatively, if only a few cases experience large increases in vote shares, the (local) average treatment effect – which is what we estimate – on vote shares may not be that large, but the impact on winning probabilities can be substantial.

The next two panels of Table 2 indicate that the (insignificant) effect of finishing first instead of second does not vary with the electoral calendar, but the second-placed advantage is five times as large in concurrent (11.2 percentage points) than in midterm (2.2 percentage points) years, with a p-value of 0.06 despite the much smaller sample size. This is consistent with theoretical expectations founded on the higher visibility and winner-takes-all nature of mayoral elections vis-à-vis midterm ones. Appendix Table A13 shows that the second-place advantage is larger in the Interior and in small municipalities, with highly significant effect sizes of 14.5 and 20.3 percentage points, respectively. The results for vote shares remain insignificant. This is consistent with expectations, but given the substantial overlap between a municipality’s location and its size – the Interior is home to 88 per cent of small municipalities but just 41 per cent of large ones; 74 per cent of the Interior’s municipalities (but only 20 per cent of the Conurbano’s) are small – we cannot determine if these results are driven by municipality size per se or by the political and demographic differences between the Conurbano and the Interior.

We also expect a larger electoral advantage of finishing second (rather than third) when the first-placed party is within reach – that is, when there is no obvious winner. Figure 5 shows that this is indeed the case: when the distance between the first- and second-placed parties is small – less than 7 percentage points – the premium of finishing second rather than third is around 50 percentage points, far more than the 9 percentage points reported in Table 2(a). Adding less competitive elections reduces this advantage almost monotonically. Figure 5(b) shows that the second-placed party receives a 3–5 percentage point boost to its vote share when it is close to the first-placed party, though these estimates are not significant.

Figure 5.

Sharp regression discontinuity (RD) estimates (points) and 95 per cent robust confidence intervals (CIs) (vertical lines) showing the effect of finishing second (rather than third) in the primary on (a) the probability of winning and (b) the vote percentage in the general election, depending on the distance between the leading and trailing party in the primary. The red horizontal lines display the RD estimates and CIs reported in Table 2(a).

Robustness

Again, calculating the running variable using all parties that contested the primary rather than just the ones surpassing the 1.5 per cent threshold does not change the results, though the reduction in power leads to mostly insignificant estimates (Table A14). Another concern is that while we have perfect balance for election-level characteristics, the parties (and candidates) that fall above or below the threshold may differ, for instance in terms of incumbency, alignment with the president (or the provincial governor), or whether they faced a competitive primary. Figure A8 documents some imbalance, especially for party ID characteristics.Footnote ¹⁸ This may be due to chance: with 112 tests, we expect 5.6 significant estimates due solely to chance, and find a total of 10. The difference is not worrisome, and in any case Table A15 shows that controlling for all variables included in the balance checks does not change the results. Using a coverage error rate (CER)-optimal instead of a MSE-optimal bandwidth (Table A16) or fitting second-order polynomials (Table A17) produces similar estimates, though the latter are much more variable. The effect of finishing first rather than second is sensitive to bandwidth choice: it begins negative at small bandwidths and then becomes zero or positive, depending on the outcome, though the estimates are never significant. In contrast, the estimate for finishing second rather than third remains pretty stable over bandwidths ranging between 5 and 35 percentage points, though the coefficients are sometimes insignificant (Figure A11).