Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
91,339 characters · 34 sections · 56 citation commands
Information and voting: Evidence from Peru's 2026 presidential election
\thispagestyle{empty}
\footnotetext[1]{Department of Mathematics, Pontificia Universidad Cat\'olica del Per\'u (PUCP); Department of Economics, University of California, Berkeley. Contact: [email removed].} \footnotetext[2]{Department of Economics, Pontificia Universidad Cat\'olica del Per\'u (PUCP); Centro de Investigaci\'on de la Universidad del Pac\'ifico (CIUP). Contact: [email removed].} \footnotetext[3]{Pontificia Universidad Cat\'olica del Per\'u (PUCP). Contact: [email removed].}
\setcounter{footnote}{0}
\noindentJEL Classifications: D72, D82, D83, C21, C93
\noindentKeywords: information design, Bayesian updating, strategic voting, exit polls, natural experiment, cardinality matching.
AI disclosure: We used Anthropic's Claude (Opus 4.7 and 4.8) to assist with prose editing, code verification, and LaTeX formatting, and Refine (refine.ink) to generate referee-style feedback on the manuscript. All substantive research decisions, analysis, and conclusions are the authors' own.
\smallPreliminary draft. All errors are our own. This paper is strictly theoretical and empirical, and does not reflect the political views or preferences of the authors.
\setcounter{page}{1}
In multi-candidate plurality elections with weak party identification, late-arriving public signals about candidate viability can change voters' minds. Peru's 2026 presidential first round provides a near-ideal setting in which to identify these effects. The election was extraordinarily fragmented, with thirty-six candidates on the ballot. On election night, the two leading polling firms (Ipsos and Datum Internacional) released their flash electoral estimates within minutes of one another, agreeing on Keiko Fujimori as the leading candidate but disagreeing on the identity of the second-place finisher. Datum named Rafael L\'opez Aliaga; Ipsos named Roberto S\'anchez. A third candidate, Jorge Nieto, was within the statistical margin of error in both estimates.
A logistical failure delivered the natural experiment. On 12 April, 187 polling tables in 13 voting centers could not be installed because the contractor failed to deliver electoral materials on time. The Jurado Nacional de Elecciones (JNE) extended voting for the affected $\approx\!55 000$ electors to Monday 13 April. These tomorrow voters cast their ballots after observing the flash estimates and the partial actas count released through the night; otherwise comparable Sunday voters did not.
This paper has three components. First, we present the political and institutional context that makes Peruvian voters unusually responsive to viability signals: a decade of executive--legislative collapse, a near-complete absence of stable party brands, and a documented voter logic of rejection rather than affiliation (the mal menor). Second, we develop a Bayesian-updating model of voting in which two pollsters issue noisy public signals about the identity of the runoff entrant. The model delivers a tractable formulation in which each voter's choice probability is a multinomial logit in which the posterior viability of a candidate is interacted with the voter's expressive match to him. Because the instrumental pull of viability is scaled by expressive affinity, the flash reallocates votes toward candidates who are both viable and ideologically acceptable to the local electorate---so a nationally viable but expressively distant candidate gains where his support is concentrated and stays flat elsewhere. The candidates whose viability the flash estimates raise are L\'opez Aliaga, S\'anchez, and Nieto. Third, the empirical section exploits the JNE ruling to identify the treated locales de votaci\'on, builds a new socioeconomic dataset from INEI Redatam census data and 2021 results, and estimates the average treatment effect on the treated by cardinality matching ZubizarretaParedesRosenbaum2014 and a within-group regression estimator.
We find that the informational shock raised the vote share of the candidates that the flash estimates rendered viable for the runoff among Monday voters relative to comparable Sunday locales, consistent with the viability-driven strategic reallocation the model predicts.
Our central contribution is empirical: exploiting the JNE ruling as a natural experiment, we identify the causal effect of overnight flash estimates on Monday voters through cardinality matching at the polling-place level (Peru's local de votaci\'on) and a within-group regression estimator, using a new dataset that characterizes these polling places socioeconomically from block-level (manzana) INEI Redatam census data merged with ONPE actas for 2021 and 2026. Adapting the single- and competing-sender frameworks of Kamenica2011 and GentzkowKamenica2017 to a multi-candidate plurality election with two public signals that may disagree, the information-design model casts the treatment as a Blackwell1953 improvement in the public experiment and thereby gives the matching estimator its interpretation as the behavioral value of that improvement (Section (ref)).
Our framework abstracts from information channels beyond the two flash estimates, such as rapid counts, television, and social-media commentary on platforms like X or Instagram, which plausibly exert their own influence on voter beliefs. This is a deliberate scope restriction rather than a claim of irrelevance; extending the analysis to these channels is left for future work.
Section (ref) sets out the political, institutional, and literature context; Section (ref) documents the 2026 empirical environment and the natural experiment; Section (ref) develops the information-design model; Section (ref) states the predictions; Section (ref) the empirical strategy; and Sections (ref)--(ref) report results and conclude.
To understand the depth of the crisis, it is necessary to trace its roots beyond 2016. Following the collapse of the Fujimori regime in November 2000, Alberto Fujimori's daughter Keiko Fujimori rebuilt the Fujimorist movement under the banner of Fuerza Popular (Popular Force). Already under Ollanta Humala (2011--2016), Fujimorist congressional blocs developed a pattern of constraining executive action, establishing legislative supremacy as a de facto norm Forbes2017. This was facilitated by the loosely worded “permanent moral incapacity” clause of Article 113 of the 1993 Constitution, which lets Congress remove the president without a threshold equivalent to high crimes, rendering the legislature more powerful than the executive and turning vacancy motions into weapons of ordinary political competition Eguiguren2017.
The 2016 elections crystallized these tensions. Pedro Pablo Kuczynski (PPK) won the runoff by fewer than fifty thousand votes over Keiko Fujimori, while Popular Force secured an absolute congressional majority---a centre-right executive facing a Fujimorist legislature with the numbers and constitutional tools to block, censure, and remove the president Mauro2026.
Entangled in the Odebrecht scandal, PPK resigned in March 2018 ahead of an imminent impeachment vote. His successor Vizcarra dissolved Congress in 2019 after a second rejected confidence vote, triggering snap elections, but was himself removed in November 2020 by a vacancy vote condemned by left-leaning sectors BarrenecheaVergara2023. Vizcarra was later sentenced to fourteen years for accepting bribes as governor of Moquegua Reuters2025Vizcarra and implicated in the Vacunagate scandal, having received COVID-19 vaccines outside the authorized trial protocol Lancet2021Vacunagate.
Protests over Vizcarra's removal forced his replacement Manuel Merino to resign after five days; Congress then appointed Francisco Sagasti to lead a transitional government to the 2021 elections, in which Pedro Castillo narrowly defeated Keiko Fujimori. Castillo's presidency was acutely fragile---nearly eighty ministerial changes in seventeen months, persistent corruption allegations, relentless obstruction---and, facing a third vacancy vote in December 2022, he attempted an unconstitutional self-coup and was removed and arrested within hours BarrenecheaVergara2023.
Vice-president Dina Boluarte assumed the presidency and aligned with the conservative bloc, triggering months of violent Andean protests that left more than sixty dead. Under her government congressional factions consolidated control over key institutions, including the Constitutional Tribunal and Public Ministry, and democratic-quality indices downgraded Peru to hybrid-regime status EIU2025; a worsening security crisis and low approval led Congress to remove her in October 2025.
Her successor Jos\'e Jer\'i lasted four months before being censured over a corruption scandal involving undisclosed meetings with Chinese businessmen. Congress then elected Jos\'e Mar\'ia Balc\'azar as its president, who assumed the presidency by constitutional succession. His appointment drew mutual accusations between Renovaci\'on Popular and Fuerza Popular, each blaming the other for enabling the ascent of a congressman from the left-wing Per\'u Libre RPP2026Balcazar, and coincided with a decline in vote intention for Rafael L\'opez Aliaga, leader of Renovaci\'on Popular Infobae2026LopezAliagaCae. This brought to nine the number of individuals to have held the office since 2016.
Therefore, as Peru approached the 2026 elections, it did so amid prolonged political instability. Human Rights Watch reports that measures passed by Congress in 2024 weakened judicial independence and constrained corruption and organized-crime investigations HRW2026Peru, in a context where most recent presidents and many legislators faced criminal allegations BarrenecheaVergara2023.
Peru represents one of Latin America's most pronounced cases of party system collapse. No major party has recently consolidated a stable and major national presence, developed genuine ideological coherence, or sustained meaningful linkages with citizens across electoral cycles. What exists instead is a succession of electoral vehicles, each organized around a candidate rather than a program, each dissolving or reconstituting itself after every election MainwaringSu2021, BarrenecheaVergara2023.
This institutional weakness is neither recent nor accidental. Its origins trace to the 1992 autogolpe, which dismantled a party system that had been gradually consolidating throughout the 1980s Tanaka2006. The 1993 Constitution, designed to reduce the role of parties as political intermediaries, produced a paradoxical outcome: not a closed system that excluded citizens, but an excessively open one that left no stable organizations between the population and power. The result was what Tanaka2006 describes as the “omnibus” phenomenon, in which national parties operate from a small nucleus in Lima with no organic territorial presence, recruiting local operators before each election in purely transactional arrangements that dissolve once voting ends Tanaka2006. Peru has consequently recorded among the highest levels of electoral volatility in Latin America MainwaringSu2021, with its leading parties changing almost entirely between electoral cycles.
The representational consequences are substantial. Each electoral cycle effectively begins from zero, with little accumulation of institutional experience and no correction of past failures, and parties cannot be held accountable for their commitments because many cease to exist before the next election. This fragility has deep historical roots: since the founding of the republic, political life has been concentrated in a Lima-based elite, and parties never developed genuine national structures or deep societal roots Chavez2022.
In the absence of stable partisan attachments, Peruvian voters have reorganized their electoral behavior around a different logic: rejection rather than affiliation Mainwaring2006. Citizens deploy their ballots not toward the candidate they support, but against the one they most fear reaching power. This phenomenon has a consistent empirical record. The negative partisanships of anti-aprismo and anti-fujimorismo remained stable across the 2011, 2016, and 2021 elections, even as the positive vehicles carrying those votes collapsed one after another Melendez2019. Humala, Kuczynski, and their successors each attracted votes defined not by full loyalty but by the imperative of blocking a feared adversary, a logic Melendez2019 terms the mal menor (lesser evil). Experimental research on political cue-taking in Peru confirms the underlying mechanism: outgroup cues consistently decrease support for policy positions, while ingroup cues produce no corresponding increase, revealing a political psychology organized more around opposition than affiliation delaCerda2025.
This is the structural context that makes strategic vote-switching not merely possible but rational. Where positive identification with any candidate is weak and provisional, a voter's first-round choice remains permanently vulnerable to revision the moment a different vote would more effectively block the feared adversary from advancing to the second round.
Peru's electoral system is governed by Ley N° 26859 and administered by three constitutionally autonomous institutions. The Jurado Nacional de Elecciones (JNE) functions as the supreme electoral tribunal, responsible for overseeing legality, resolving disputes, and proclaiming official results; its decisions in electoral matters are final and unappealable. The Oficina Nacional de Procesos Electorales (ONPE) bears operational responsibility for organizing and executing elections: it deploys decentralized offices throughout the country, appoints polling-station coordinators through public competition, and coordinates with the police and armed forces to guarantee order on election day. A third body, the Registro Nacional de Identificaci\'on y Estado Civil (RENIEC), maintains the voter registry and issues the Documento Nacional de Identidad (DNI), without which no citizen may cast a vote.
General elections are held every five years on the second Sunday of April (first round), and voting is mandatory for citizens between the ages of 18 and 70 (optional thereafter); failure to vote is sanctioned by a monetary fine and entails temporary administrative restrictions until the obligation is regularized. Presidential candidates must win an outright majority of valid votes; if none does, a runoff is held within thirty days between the two leading candidates. Congressional elections take place simultaneously. The 2026 election marked a structurally significant change: for the first time since 1990, voters elected a bicameral Congress comprising 130 deputies and 60 senators. Deputies, as the lower chamber, are tasked with presenting legislation and overseeing the Cabinet, while senators, as the upper chamber, review bills, approve key institutional appointments, and hold the power to remove the president. Access to seats in either chamber is conditioned on an electoral threshold (valla electoral) that a party must clear on two counts simultaneously: it must obtain at least five percent of the valid votes cast nationwide for that chamber, and it must be allotted, in the seat-distribution procedure, a number of seats equal to at least five percent of the chamber's legal membership---seven of the 130 deputies or three of the 60 senators. This dual requirement is designed to limit fragmentation in a historically atomized party system, though it carries the trade-off of potentially reducing effective representation when a significant share of the electorate votes for parties that fail to clear the threshold.
Voting is organized at the level of the individual mesa de sufragio (polling table), where ballots are cast, counted, and recorded in official actas electorales transmitted to ONPE's central tally. The polling table is therefore the foundational unit of the electoral chain, and any disruption at this level---such as a failure to install a table on time---directly interrupts the exercise of suffrage.
Our paper draws on four literatures. The first is information design and Bayesian updating Kamenica2011, Bergemann2019, which studies how the signals available to a decision-maker can be structured; nearer our setting are Alonso2016, Chan2019, and---closest, with Datum and Ipsos issuing potentially conflicting calls--- GentzkowKamenica2017 on competing senders. We borrow its apparatus (signal structure, Bayesian updating, and the Blackwell1953 comparison of experiments) but not its premise of a strategic sender: we neither model pollsters as designing estimates to sway voters nor take any stance on their intent. What the framework supplies instead are the definitions and structure for our object of interest---the natural experiment assigns Sunday and Monday voters to two experiments over the identity of the runoff entrant, and we estimate the behavioral value of the Blackwell improvement from the coarser (pre-flash) to the finer (post-flash) experiment (Section (ref)).
To take this structure to behavior, we turn to a second literature on choice modeling. The random-utility multinomial logit Train2009 microfounds the model, mapping the expressive and viability components into individual choice probabilities. Because the ballot is secret, however, we never observe individual choices: our outcomes are aggregate returns at the acta level. The logit therefore plays a structural role, disciplining the aggregate vote shares we observe rather than being fit to individual data.
The methodological core of the paper belongs to a third literature, on matching for causal inference. We use cardinality matching on pre-treatment covariates ZubizarretaParedesRosenbaum2014, ViscontiZubizarreta2018, which recovers the ATT on treated locations under selection on observables, estimated with within-group difference-in-means and regression estimators PageLenardKeele2020.
Finally, two empirical literatures provide background rather than method. The first studies media and political updating DellaVigna2007, Enikolopov2011, Gentzkow2014; its closest antecedent to our design is Morton2015, who exploits a French natural experiment to identify exit-poll-induced turnout and bandwagon effects via the geographic and temporal staggering of exposure to a public signal. The second concerns voting in weakly institutionalised democracies, where Lupu2016, Dargent2015, Vergara2018, and Levitsky2022 ground the strategic voting under the mal menor logic of Section (ref).
The 2026 ballot featured 36 candidates spanning the full ideological range (Table (ref); TuVotoPeru2026), and the field was exceptionally fragmented. Fujimori (Fuerza Popular) led on a free-market, mano dura, socially conservative platform, but no other candidate pulled clear: a pack of contenders competed closely for the second runoff slot, spanning the spectrum. On the right, L\'opez Aliaga (Renovaci\'on Popular) carried the most explicitly hard-right discourse on security, religion, and social values; on the left, S\'anchez (Juntos por el Per\'u) ran a consistently left-wing platform---protectionist, statist, and progressive; and Belmont, \'Alvarez, and Nieto held intermediate or off-axis positions, Nieto in particular running as a “fresh option” outside the traditional left--right divide, with L\'opez Chau also among the leading group. Given this ideological spread but weak programmatic differentiation among the front-runners, voter choice is more plausibly driven by perceived viability than by policy, consistent with the mal menor logic of Section (ref).
In the polls, three features defined the pre-electoral environment. Vote intention was historically low (Figure (ref)): Fujimori ran between roughly $14\%$ and $19\%$ in the final month, while the candidates competing for second place clustered in a tight 7--13% band, with five or six within a few points of one another. The race was volatile, with rankings shifting substantially between waves---Ipsos and Datum ran the two longest regular series (Figures (ref) and (ref)). And the pollsters disagreed on who occupied second place, a disagreement that persisted into their election-night flash estimates (Section (ref)). The six official JNE debates appear to have moved intentions, with \'Alvarez and Belmont gaining and L\'opez Aliaga declining ipsos2026preelect.\footnote{Series compiled from the La Encerrona tracker, \url{https://laencerrona.pe/2026/02/10/encuestas}; Peruvian law prohibits poll publication in the week before the election.}
The first round produced no clear front-runner: the four leading candidates finished within six points (Fujimori $17.06\%$, S\'anchez $12.04\%$, L\'opez Aliaga $11.90\%$, Nieto $11.03\%$), in a geographically fragmented result---an electoral “archipelago” splitting Lima and the north coast between the two right-of-centre candidates, the Andes toward S\'anchez, and the urban south toward Nieto Martinelli2026archipielago. With barely a point separating the second runoff slot, its occupant was unresolved on election night.
At 18:00 on 12 April, Ipsos and Datum released their flash electoral estimates (Table (ref)). Both placed Fujimori first but disagreed on second: Datum put L\'opez Aliaga second ($12.8\%$); Ipsos put S\'anchez second ($12.1\%$) with Belmont third ($11.8\%$). The trajectory from the three Ipsos simulacros to the flash showed S\'anchez and Belmont surging and \'Alvarez declining (Figure (ref)). The estimates thus confirmed Fujimori's passage while leaving the second slot open---a potential right-versus-right runoff, but with S\'anchez registering as a viable left-of-center contender in Ipsos and the centrist Nieto within the margin of error in both estimates, so that the candidates in contention for the single remaining slot spanned the full spectrum: L\'opez Aliaga on the radical right, Nieto at the center, and S\'anchez on the left.
This section develops the framework that organizes the empirical analysis. Its premise is that the Ipsos and Datum flash estimates are public signals about the identity of the runoff entrant, and that such a signal can in principle shift a voter's choice by revising her beliefs about which candidates are viable. We microfound this channel at the level of the individual voter. The objective is operational: a closed-form expression for the probability that voter $i$ chooses candidate $c$ after observing the two flash estimates, whose within-table aggregate is the effect our matching estimator identifies. We first model each flash estimate as a noisy public signal and derive the voter's posterior over the runoff entrant; we then embed that posterior in a random-utility model of individual choice in which the instrumental value of a viable candidate is scaled by the voter's expressive match to him.
The state space $\Theta$ indexes the identity of the second-place finisher. Fujimori's lead in first place is uncontested across all major pre-electoral polls and in both flash estimates, so uncertainty is confined to who advances with her to the runoff. Guided by the flash data of Section (ref)---in which Datum names L\'opez Aliaga, Ipsos names S\'anchez, and Nieto polls within a percentage point of the named candidates in both estimates---we restrict attention to the three candidates the flash estimates render viable:
A belief is a probability distribution on $\Theta$; the set of beliefs is the unit simplex $$ \Delta(\Theta) = \left\{\mu \in \mathbb{R}_+^{K} : \sum_{\theta} \mu(\theta) = 1\right\}, $$ so that $\mu_j \in \Delta(\Theta)$ for every voter $j$. The prior $\mu_0 \in \Delta(\Theta)$ is the belief held before any signal is observed.
A signal---or experiment\footnote{An experiment in the sense of Blackwell1953 is a pair consisting of a finite signal space and a family of probability distributions over it, one for each state of the world.}---is a pair $(S, \pi)$ consisting of a finite signal space $S$ and a Markov kernel $\pi : \Theta \to \Delta(S)$.
We model each pollster's flash estimate as a symmetric $\eta$-noisy experiment: $S = \Theta$ and
Nature draws $\theta$; the experiment announces it correctly with probability $1 - \eta$ or, with probability $\eta$, announces one of the $K-1$ wrong states uniformly. The single number $\eta$ absorbs both sampling and non-sampling error in a way that admits a clean Blackwell ordering: $\pi^{\eta_1} \succeq_B \pi^{\eta_2}$ whenever $\eta_1 \le \eta_2$ Blackwell1953.
The 2026 flash produced calls $(s_D^*, s_I^*) = (\text{RLA}, \text{RS})$. We assume the two signals are conditionally independent given the state.
The assumption is supported by the operational independence of the two firms' samples and weighting procedures; we discuss it further in Section (ref).
Candidates are indexed by $c \in \mathcal{C}$ (the full 36-candidate ballot of Section (ref), of which $\Theta$ is the viable subset for second place); voter $j$ has type $t_j = (\alpha_j, \beta_j, M_j)$. The match score $M_j(c) \in [0,1]$ summarizes voter $j$'s ideological and identity-based affinity with candidate $c$; empirically it is proxied at the location level by the right-of-centre orientation of the polling place (Section (ref)).
Let $\mathrm{runoff}(\theta) \triangleq \{\text{Fujimori}, \theta\}$ denote the pair of candidates that advance to the runoff in state $\theta \in \Theta$. Voter $j$'s utility from voting for $c$ in state $\theta$ is
with $\alpha_j > 0$, $\beta_j \geq 0$, and i.i.d.\ Type-I Extreme Value taste shocks $\varepsilon_{jc}$. The instrumental term is interacted with the expressive match $M_j(c)$: learning that $c$ is viable moves voter $j$ toward him only if $M_j(c)$ is appreciable, so a viable but expressively rejected candidate ($M_j(c) \approx 0$) exerts no instrumental pull. This multiplicative structure lets a nationally viable candidate gain votes only where he is also expressively acceptable.
Define the perceived-viability function $q_c : \Delta(\Theta) \to [0, 1]$ that maps a posterior belief over runoff entrants into the perceived probability that $c$ is in the runoff:
Under the standard Gumbel assumption on the idiosyncratic utility shocks Train2009, voter $j$'s probability of choosing candidate $c$ after observing the signal pair $s \in S \times S$ admits the closed-form logit expression
where $\mu_j^{s} \coloneqq \mu_j(\,\cdot \mid s\,) \in \Delta(\Theta)$ is the posterior of Proposition (ref). The signal enters (ref) only through the viability terms $q_c(\mu_j^{s})$, but its impact on $c$'s choice probability is scaled by the expressive match $M_j(c)$: the same posterior shift moves a voter toward $c$ in proportion to how acceptable $c$ already is to her. Because the 2026 calls are discordant ($s_D^* \neq s_I^*$), the posterior of Proposition (ref) splits viability mass between $s_D^*$ and $s_I^*$ rather than concentrating it on a single candidate: each named candidate gains in viability, but by strictly less than under a concordant call.
The behavioral content of (ref) is a viability--affinity interaction: a rise in $q_c(\mu_j^{s})$ raises $\mathbb{P}_j(c \mid s)$ by an amount increasing in $M_j(c)$. The more a voter already favors $c$, the more a gain in his viability moves her toward him; where affinity is low, the same gain barely shifts her. The effect on L\'opez Aliaga and S\'anchez is therefore symmetric: each gains in proportion to local affinity---L\'opez Aliaga where the electorate leans right, S\'anchez where it leans left---and neither draws voters from the other side. The mal menor logic of Section (ref) reallocates support along existing affinities, not across them.
The structural model of Section (ref) operates on the individual voter $j$, who casts a vote with logit probability (ref). The data are silent at this level: the ONPE actas record only vote counts at each polling table, and the secret ballot precludes any link between voters and votes. The empirical strategy must therefore identify the model-implied effect from table-level shares alone, which this subsection makes precise.
Voting locations are indexed by $\ell$ and actas by $a$, with $a \in \ell$ denoting that acta $a$ is recorded at location $\ell$. Let $T_\ell \in \{0,1\}$ be the treatment indicator, equal to $1$ if location $\ell$ voted on Monday 13 April 2026---after the flash estimates were broadcast---and $0$ otherwise; each acta inherits its location's status, so $T_a \equiv T_\ell$ for $a \in \ell$. The treated and control locations are \[ \mathcal{T} \;\coloneqq\; \left\{\, \ell : T_\ell = 1 \,\right\} \;=\; \left\{\, \ell_1, \ldots, \ell_{13} \,\right\}, \qquad \mathcal{C}_0 \;\coloneqq\; \left\{\, \ell : T_\ell = 0 \,\right\}, \quad \left| \mathcal{C}_0 \right| \approx 340. \] For the $i$-th treated location ($i = 1, \ldots, 13$), let $A_{\ell_i} = \left\{\, a_{\ell_i}^{1}, \ldots, a_{\ell_i}^{n_i} \,\right\}$ be its set of actas, with $n_i = \left| A_{\ell_i} \right|$. The total number of treated actas is \[ N_T \;\coloneqq\; \sum_{i=1}^{13} n_i \;=\; 187. \]
At acta $a$ with $N_a$ valid votes, let $V_{jc} = \mathbf{1}\{j \text{ votes for } c\}$ for voter $j \in a$ and candidate $c$. The observed share is \[ Y_{ac} \;=\; \frac{1}{N_a} \sum_{j \in a} V_{jc}. \] Under (ref), $\mathbb{E}[V_{jc} \mid \mu_j] = \mathbb{P}_j(c \mid \mu_j)$. Write $\mu_j^{(0)}$ for voter $j$'s baseline posterior and $\mu_j^{(1)} \equiv \mu_j(\,\cdot \mid s_D^*, s_I^*)$ for her posterior after observing the two flash signals (Proposition (ref)). The expected potential outcome of acta $a$ under treatment $d \in \{0, 1\}$ is \[ \mathbb{E}[Y_{ac}(d)] \;=\; \frac{1}{N_a} \sum_{j \in a} \mathbb{P}_j\left(c \mid \mu_j^{(d)}\right). \] Let \[ \delta_{jc} \;\coloneqq\; \mathbb{P}_j\left(c \mid \mu_j^{(1)}\right) - \mathbb{P}_j\bigl(c \mid \mu_j^{(0)}\bigr) \] denote the individual information effect for voter $j$ on candidate $c$. Differencing the potential outcomes yields the acta-level treatment effect as the within-acta average of the $\delta_{jc}$:
The $\delta_{jc}$ are not observable; their within-acta average $\tau_{ac}$ is the smallest object the model speaks to that is also visible in the data.
The estimand is the acta-weighted average of $\tau_{ac}$ across the treated population,
which is the average treatment effect on the treated (ATT) for candidate $c$. Its model-implied counterpart is
where $F_a$ is the within-acta empirical distribution of voter types (the probability measure that places mass $1/N_a$ on each voter $j \in a$, so that $$ \int g(j)\, dF_a(j) = N_a^{-1} \sum_{j \in a} g(j) $$ for any function $g$ of voter characteristics). Equation (ref) is the population analogue of the table-level effect computable in closed form from the structural model; equation (ref) is its causal counterpart in the potential-outcomes framework. The two are linked by the identification result below.
The institutional ground for Assumption (ref) is the contractor's logistical protocol of 12--13 April 2026: which locations were displaced was decided on operational grounds (transport routes, warehouse-to-centre delivery times, the operational profile of the districts served) that are correlated with district-level socioeconomic conditions but orthogonal, by design, to the political preferences of the location's electorate. No political variable enters the contractor's protocol. Conditional on the socioeconomic and demographic covariates that drive the operational protocol, the residual variation in $T_\ell$ is therefore as good as random.
Assumption (ref) rules out between-location spillovers.
Two readings of (ref) matter for what follows. The identity $\tau_c = \Delta_c$ ties the empirical target to the model and licenses the sign predictions of Section (ref), while the difference between the treated mean of $Y_{ac}$ and the control-based counterfactual is the estimable object the matching procedure of Section (ref) constructs.
Proposition (ref) identifies $\tau_c$ as a causal ATT but not why it is an information effect. The reason: treatment is a change of Blackwell1953 experiment over the state space $\Theta$ of Section (ref). A Sunday voter ($T_\ell = 0$) decides before the flash, under the null experiment $\pi^{\varnothing}$, whose single signal leaves the posterior at the prior, $\mu(\,\cdot \mid \pi^{\varnothing}) = \mu_0$. A Monday voter ($T_\ell = 1$) decides after it, under the flash experiment
the conditionally independent (Assumption (ref)) combination of the Datum and Ipsos kernels, with realized value $(s_D^*, s_I^*) = (\mathrm{RLA}, \mathrm{RS})$ inducing the posterior $\mu_i^{(1)}$ of Proposition (ref). The null discards its signal---it is a garbling of every experiment---so any experiment Blackwell-dominates it; hence $\pi^{\mathrm{F}} \;\succeq_B\; \pi^{\varnothing}$, strictly once at least one flash kernel is informative, i.e. $\eta_D, \eta_I < 1 - 1/K$. Treatment thus moves voters, comparable in $H_a$ by Assumption (ref), from $\pi^{\varnothing}$ to $\pi^{\mathrm{F}}$, and $\tau_c$ is the population behavioral value of that improvement. Three results stack into
whose right-hand side Section (ref) estimates.
The model delivers five sign predictions for the acta-level treatment effects $\tau_c$ defined in (ref), all stated for the treated sample, which is composed of polling stations located in Lima Sur. The flash raises perceived viability $q_c(\mu_j^s)$ above the prior for three candidates---Nieto (within a percentage point of both named candidates in either estimate), López Aliaga (Datum's named candidate), and Sánchez (Ipsos's named candidate)---confirms Fujimori in every state, and leaves the rest non-viable. The basic prediction is that the viable candidates gain and the non-viable candidates lose.
By the viability--affinity interaction in (ref), a rise in $q_c(\mu_j^s)$ moves voter $j$ toward $c$ in proportion to his expressive match $M_j(c)$, so within the right-leaning Lima sample the gains rank by match, not by viability alone. Nieto---off-axis, less extreme, a clean alternative to both blocks---has the highest match; López Aliaga is right-aligned but tied to the discredited congressional pact Infobae2026LopezAliagaCae, which lowers his; and Sánchez, left-anchored with a provincial base Martinelli2026archipielago, is expressively distant in Lima. The viable gains therefore rank theoretically Nieto $>$ López Aliaga $>$ Sánchez.
The aggregation constraint $\sum_{c \in \mathcal{C}} Y_{ac} = 1$ implies $\sum_{c \in \mathcal{C}} \tau_c = 0$, so the nonnegative effects in H(ref)--H(ref) are offset by the losses in H(ref); the leading candidates need not net to zero individually, since the remainder falls on the long tail of minor candidates.
Equation (ref) reduces identification to constructing a credible counterfactual mean for the treated actas. With only thirteen treated locations, a propensity-score model is fit on too few treated units to be reliable---the estimated scores are noisy and produce extreme weights---and treatment varies only at the location level. We therefore match at that level: cardinality matching selects the largest set of control locations whose average covariates equal those of the treated locations, up to preset tolerances, on the covariate vector $H_a$ ZubizarretaParedesRosenbaum2014, ViscontiZubizarreta2018. Unlike the propensity score, it imposes covariate balance directly, as an explicit constraint, rather than relying on a fitted treatment-probability model. On the matched sample we estimate $\tau_c$ two ways: a within-group difference-in-means estimator, and a within-group regression estimator that additionally corrects for any covariate differences that remain between treated and control actas within a matched group PageLenardKeele2020.
The analysis sample covers Lima Sur and comprises 354 voting locations (locales de votaci\'on) and 4930 actas. Of these locations, thirteen are treated and 341 belong to the control pool. The dataset is built from two primary sources: the Instituto Nacional de Estad\'istica e Inform\'atica (INEI) and the Oficina Nacional de Procesos Electorales (ONPE).
The first source is INEI, which provides georeferenced census information at the block (manzana) level. These data are used to characterize the socioeconomic and demographic conditions of each voting location's surroundings, including total population, recent in-migrants, linguistic composition, private health-insurance coverage, educational attainment, literacy, ethnic self-identification, and indicators of access to basic services and housing conditions.
The second source is ONPE, which provides the electoral results of the 2021 and 2026 general elections, as well as information from the 2026 electoral roll. These data supply the electoral outcomes, the prior-election variables used to characterize historical voting behavior, and the demographic composition of registered voters at each location.
Census variables are linked to each voting location through a spatial procedure. For each voting location $\ell$, we construct a 1{,}000-meter (1\,km) area of influence---a buffer---around the polling place and identify the census blocks that fall within it. Let $B(\ell)$ denote the set of census blocks within this buffer and let $\mathrm{pop}_b$ be the population of block $b$. For any block-level census variable $v_b$, the corresponding location-level measure is computed as the population-weighted average \[ V_\ell \;=\; \frac{\sum_{b \in B(\ell)} \mathrm{pop}_b\, v_b} {\sum_{b \in B(\ell)} \mathrm{pop}_b}. \] The resulting variables describe the population residing in the immediate surroundings of each voting location rather than the average conditions of the administrative district.
This spatial criterion is also consistent with the institutional logic of voter-location assignment in Peru. Through the Elige tu local program, voters are allowed to select up to three preferred polling locations, typically based on proximity to their residence, and to rank them according to their desired priority. The final assignment depends on the availability of capacity at the selected locations, but the mechanism nevertheless creates a direct link between voting locations and the residential areas surrounding them.
The 2021 electoral results are then linked to the voting locations used in the analysis. Where a voting location can be tracked across electoral processes, the correspondence is made directly. Where no exact match exists, the assignment relies on geographic proximity and similarity in the number of polling tables (mesas de votaci\'on), so as to identify the most comparable prior-election location.
Matching uses nine location-level covariates, all measured before 13 April 2026 and constant across a location's actas. We index these covariates by the acta $a$ under the convention $H_a = H_{\ell(a)}$, since the balance constraints (ref) act on acta-weighted means. The covariates fall into three blocks.
The first block contains four socioeconomic variables constructed from INEI block-level census data using the spatial procedure described above: \[ V \in \{\mathrm{SES}, \mathrm{Indig}, \mathrm{EDU}, \mathrm{Net}\}. \] These variables are the socioeconomic stratum $\mathrm{SES}_\ell$, measured by the INEI stratum index from $1$ lowest to $5$ highest; the indigenous-mother-tongue share $\mathrm{Indig}_\ell$; the completed-higher-education share $\mathrm{EDU}_\ell$; and the internet-access share $\mathrm{Net}_\ell$. The last three variables lie in $[0,1]$.
The second block summarizes prior political orientation using 2021 first-round vote shares. Let $s_{\ell,c}^{2021}$ be candidate $c$'s share of the valid vote at location $\ell$ in the first round of the 2021 general election, and let $\mathcal{L}^{2021}$, $\mathcal{R}^{2021}$, and $\mathcal{F}^{2021}$ be the disjoint sets \[ \mathcal{L}^{2021} = \{\text{Per\'u Libre, Juntos por el Per\'u, Partido Morado, Acci\'on Popular}\}, \] \[ \mathcal{R}^{2021} = \{\text{Avanza Pa\'is, Renovaci\'on Popular}\}, \qquad \mathcal{F}^{2021} = \{\text{Fuerza Popular}\}. \] The three political-orientation variables are \[ \mathrm{Left}_\ell^{2021} = \sum_{c \in \mathcal{L}^{2021}} s_{\ell,c}^{2021}, \qquad \mathrm{Right}_\ell^{2021} = \sum_{c \in \mathcal{R}^{2021}} s_{\ell,c}^{2021}, \qquad \mathrm{Fuji}_\ell^{2021} = \sum_{c \in \mathcal{F}^{2021}} s_{\ell,c}^{2021}. \] All three variables lie in $[0,1]$. Splitting prior orientation into three blocs rather than imposing one balance constraint per 2021 candidate is required by the narrow treated sample, which cannot support a separate balance constraint for each candidate under (ref). Because the three sets are disjoint, the shares do not double-count votes.
The third block comes from the 2026 electoral roll at the location: the share of female registered voters $\mathrm{Fem}_\ell \in [0,1]$ and the mean registered-voter age $\mathrm{Age}_\ell$, standardized across locations to have mean $0$ and unit variance. The resulting balance vector is
Matching is conducted within district. The district marginal is imposed on (ref) as a fine-balance side constraint. The cardinality step selects $\mathcal{C}^\star$ from the initial control pool $\mathcal{C}_0$, and the second stage assigns each selected control to a within-district treated location by Mahalanobis distance, forming the matched groups $G_i$ described in Section (ref).
Table (ref) reports the raw, pre-matching distribution of the balance covariates and the 2026 outcomes for the thirteen treated and the 341 control locations. Moments are acta-weighted, so each location enters in proportion to its number of actas, consistent with the acta-weighted estimand $\tau_c$ in (ref). The reported $N$ counts voting locations, which is the level at which treatment is assigned and at which the covariates are constant.
Two patterns stand out. First, treated locations sit below the control pool on the socioeconomic gradient: they have a lower socioeconomic stratum ($2.21$ vs.\ $2.60$), lower completed-higher-education share ($0.249$ vs.\ $0.312$), lower internet access ($0.361$ vs.\ $0.453$), and a higher indigenous-language share ($0.110$ vs.\ $0.091$). This is precisely the imbalance that the matching procedure in Section (ref) is designed to remove. Second, political-orientation covariates are already relatively close before matching: the 2021 left, right, and Fujimori shares differ by at most 2.2 percentage points, and the female-voter share is nearly identical across groups ($0.492$ vs.\ $0.499$). This pattern is consistent with Assumption (ref): displacement is related to the socioeconomic profile of the served areas, rather than to their prior electoral preferences.
The treated group is also tighter, with smaller standard deviations on every covariate, and its support lies inside the control range throughout. Thus, common support holds before matching and the matched controls are drawn by interpolation rather than extrapolation.
Panel B reports the seven leading candidates' 2026 first-round vote shares, the outcomes $Y_{ac}$. The role of these gaps differs from that of the covariate gaps in Panel A. Covariate imbalance is a nuisance that the matching design removes, whereas outcome gaps are the quantities of interest. If the overnight Ipsos and Datum estimates affected Monday ballots, treated--control differences in vote shares are precisely what the design should detect.
The raw outcome contrasts are sizable. Treated locations show higher mean shares for Fujimori ($0.236$ vs.\ $0.201$), S\'anchez ($0.064$ vs.\ $0.031$), L\'opez Aliaga ($0.186$ vs.\ $0.161$), and especially Nieto ($0.311$ vs.\ $0.148$). They show lower shares for Belmont ($0.080$ vs.\ $0.098$), \'Alvarez ($0.042$ vs.\ $0.106$), and L\'opez Chau ($0.015$ vs.\ $0.060$). These raw contrasts, however, do not by themselves identify the treatment effect because they still combine the Monday treatment with the socioeconomic imbalance described in Panel A. The matched design of Section (ref) isolates the treatment component, with estimates and interpretation reported in Section (ref).
The empirical procedure returns, in two stages (Section (ref)), a control sample $\mathcal{C}^\star \subseteq \mathcal{C}_0$ and a partition of $\mathcal{T} \cup \mathcal{C}^\star$ into thirteen matched groups \[ G_i \;=\; \left\{ \ell_i,\, \tilde\ell_1^{\,i},\, \tilde\ell_2^{\,i},\, \ldots,\, \tilde\ell_{p_i}^{\,i} \right\}, \qquad i = 1, \ldots, 13, \] where $\tilde\ell_r^{\,i} \in \mathcal{C}^\star$ is the $r$-th control location matched to the treated location $\ell_i$, and $p_i$ is the number of controls in group $G_i$. Group sizes are not fixed: $p_i$ is determined endogenously by the matching and constrained to lie between three and ten, $3 \le p_i \le 10$. The lower bound rules out groups so small that a single atypical control would drive the within-group contrast; the upper bound keeps a treated stratum from being diluted by a large, internally heterogeneous control set and preserves the locality of the comparison. Inside group $G_i$, the actas of the $r$-th control are \[ A_{\tilde\ell_r^{\,i}} \;=\; \left\{ a_{\tilde\ell_r^{\,i}}^{\,1}, \ldots, a_{\tilde\ell_r^{\,i}}^{\,n_{\tilde\ell_r^{\,i}}} \right\}, \] and the total number of control actas in $G_i$ is $N_i^C \coloneqq \sum_{r=1}^{p_i} n_{\tilde\ell_r^{\,i}}$. Aggregating across groups, $N_C \coloneqq \sum_{i=1}^{13} N_i^C$ is the size of the matched control sample at the acta level. The outcome at acta $a \in A_{\ell_i} \cup \bigcup_{r} A_{\tilde\ell_r^{\,i}}$ is the share $Y_{ac}$ for candidate $c$. Treatment is constant within group: every acta in $A_{\ell_i}$ has $T_\ell = 1$ and every acta in $\bigcup_{r} A_{\tilde\ell_r^{\,i}}$ has $T_\ell = 0$. Each treated location therefore defines its own matched stratum, and identification within $G_i$ is the comparison between treated \emph{actas} at $\ell_i$ and the pooled \emph{actas} of its $p_i$ matched controls. Finally, $n_i \equiv n_{\ell_i} = |A_{\ell_i}|$.
Cardinality matching selects the largest control sample that satisfies pre-specified balance constraints, rather than minimizing a pairwise distance ZubizarretaParedesRosenbaum2014, ViscontiZubizarreta2018. With only thirteen treated clusters, propensity-score models are unstable; fixing the balance target ex ante and maximizing the matched sample subject to it is more robust. Let $q_\ell \in \{0,1\}$ indicate inclusion of control location $\ell \in \mathcal{C}_0$. The program maximizes the matched control sample, measured in actas, subject to acta-weighted balance on every component of $H_a$:
where $J = \dim(H_a)$, $s_j$ is the pooled standard deviation of $H_{j,a}$, and the left-hand side is the absolute standardized mean difference (SMD) of covariate $j$ between the treated sample and the selected controls. The acta-weighted means are \[ \bar H_{j,T} = \frac{1}{N_T} \sum_{i=1}^{13} \sum_{a \in A_{\ell_i}} H_{j,a}, \qquad \bar H_{j,\mathcal{C}^\star} = \frac{1}{N_{\mathcal{C}^\star}} \sum_{\ell \in \mathcal{C}_0} q_\ell \sum_{a \in A_\ell} H_{j,a}, \] with $N_{\mathcal{C}^\star} = \sum_{\ell} q_\ell n_\ell$ and selected controls $\mathcal{C}^\star = \{\ell : q_\ell = 1\}$. The baseline imposes a strict balance rule: every covariate must meet $\delta_j = 0.10$, that is $|\mathrm{SMD}_j| \le 0.10$ for all $j$ with no exceptions. The $0.10$ ceiling is the conventional threshold below which covariate imbalance is treated as negligible Austin2009, Stuart2010; entering it as a hard, per-covariate constraint---rather than as an average or an aggregate target---ensures no single covariate is left imbalanced in exchange for tighter balance elsewhere. Section (ref) reruns the match at the tighter and looser caps $\delta_j \in \{0.05, 0.15\}$ to confirm the result is not an artifact of the $0.10$ cutoff. Fine balance on the district marginal enters as a side constraint: a treated district must contribute the same share of selected controls as of treated locations. Problem (ref) is solved in Python.
Problem (ref) selects which control locations enter the sample, but not which of them serves as a comparison for each treated location. The estimators below need that assignment: each treated location must sit in a group with its own comparable controls. A second stage therefore partitions $\mathcal{T} \cup \mathcal{C}^\star$ into the thirteen groups $G_i$, one per treated location, with each treated location receiving between three and ten controls. Controls are assigned by Mahalanobis distance on the continuous sub-vector $Z_\ell^{\mathrm{cont}} :=(\mathrm{SES}_\ell,\; \mathrm{EDU}_\ell,\; \mathrm{Left}_\ell^{2021},\; \mathrm{Right}_\ell^{2021},\; \mathrm{Fuji}_\ell^{2021},\; \mathrm{Fem}_\ell,\; \mathrm{Age}_\ell)$, \[ d(\ell, \ell') = (Z_\ell^{\mathrm{cont}} - Z_{\ell'}^{\mathrm{cont}})^\top \Sigma_{Z^{\mathrm{cont}}}^{-1} (Z_\ell^{\mathrm{cont}} - Z_{\ell'}^{\mathrm{cont}}), \] with $\Sigma_{Z^{\mathrm{cont}}}$ the empirical covariance of $Z_\ell^{\mathrm{cont}}$, so that a small $d$ marks two very similar locations. A control may be assigned to a treated location only if it lies inside a Mahalanobis caliper, $d(\ell, \ell') \le 3$; controls beyond it are left unassigned rather than forced into a poorly matched group. Controls failing the caliper are dropped from the matched sample, so the balance reported in Section (ref) is computed on the retained controls; otherwise the stage only groups the already-balanced controls into locally homogeneous comparison sets.
We estimate $\tau_c$ on the matched groups $\{G_i\}_{i=1}^{13}$ with two estimators that share this design but differ in how they handle residual within-group imbalance. The first is a nonparametric difference-in-means (DM) estimator. The treated and control acta-weighted means in $G_i$ are
with group contrast $\hat\Delta_{i,c} \coloneqq \bar Y_{i,c}^{\,T} - \bar Y_{i,c}^{\,C}$. The ATT estimator weights each group by its acta share of the treated sample, $w_i \coloneqq n_i / N_T$:
These weights reproduce the acta-weighted ATT of (ref): a treated location with more actas counts proportionally more. Equation (ref) imposes no functional form linking $Y_{ac}$ to $H_a$; it relies entirely on the balance from (ref). As a robustness check, Section (ref) also reports the location-weighted variant (superscript $L$), which weights each treated location equally regardless of its number of actas:
The second estimator adjusts for residual imbalance in the acta-level demographics $X_a$ (age and female participation), following standard practice for clustered observational studies PageLenardKeele2020. Group by group, run
where $T_\ell = 1$ for $a \in A_{\ell_i}$ and $0$ otherwise, and $\psi_{i,c}$ is the within-group slope vector on $X_a$.
The location covariates $Z_\ell$ are omitted. They are constant on the treated arm (one treated location per group), but on the control arm $Z_\ell$ varies across the matched controls and is therefore not swept out by the group fixed effect; what makes its omission innocuous is the balance constraint (ref), which holds each treated location's $Z_\ell$ to its matched-control mean within the imposed tolerance, so the residual imbalance in $Z_\ell$ is minimized by design rather than identically absorbed by the fixed effects. The group-by-group aggregate of the within-group treatment effects, with the same weights,
uses the group-specific slopes $\psi_{i,c}$ of (ref). We instead compute the more parsimonious pooled, acta-weighted regression
with a matched-group fixed effect $\alpha_{g(\ell), c}$ for the group containing $\ell$, a single treatment coefficient $\tau_c$, and slopes $\psi_c$ common across groups, each acta weighted so that the treated and control arms of $G_i$ each carry total weight $w_i$ (each treated acta gets $1/N_T$; each control acta in $G_i$ gets $w_i/N_i^C$); we set $\widehat\tau_c^{\,\mathrm{R}}$ equal to the OLS estimate of $\tau_c$ in (ref). Dropping $X_a$ recovers $\widehat\tau_c^{\,\mathrm{DM}}$ exactly. The two regression adjustments coincide only when the covariate slopes are homogeneous across groups ($\psi_{i,c} = \psi_c$) or $X_a \perp T_\ell$ within every $G_i$; with heterogeneous slopes (ref) and (ref) differ, because the latter pools the covariate adjustment across groups while the former does not.
The two estimators differ only in how they treat residual within-group imbalance: $\widehat\tau_c^{\,\mathrm{DM}}$ takes the raw contrast, $\widehat\tau_c^{\,\mathrm{R}}$ adjusts for $X_a$ through the common slopes $\psi_c$ of (ref). They coincide when $X_a \perp T_\ell$ within $G_i$, and their gap measures how much the adjustment matters. We report $\widehat\tau_c^{\,\mathrm{DM}}$ as the headline estimate and $\widehat\tau_c^{\,\mathrm{R}}$ as the regression-adjusted check.
The design's validity rests on the matched controls reproducing the treated covariate distribution. The main diagnostic is the post-match SMD of each component of $H_a$,
with the practical threshold $|\mathrm{SMD}_j| < 0.10$; Table (ref) reports the before- and after-matching SMDs. The full battery---per-covariate Kolmogorov--Smirnov statistics, before-and-after density plots of $H_{j,a}$, the geographic spread of treated and selected control locations, and the matched-sample sizes $|\mathcal{C}^\star|$ and $N_C$ (which show whether the comparison rests on a wide control pool or a few influential units)---is reported in Appendix (ref).
Given the small treated sample, the estimate is stress-tested four ways.
The main limitation is the small number of treated locations: although the control pool is large, the independent causal information comes from $L_T = 13$ clusters, so the estimates are evidence for the displaced locations rather than a generalizable nationwide effect. Second, the match balances only observed covariates, so sensitivity to unobserved confounding remains a concern the design cannot rule out.
The baseline matched design retains the thirteen treated locations (187 actas) and selects 178 control locations ($2386$ actas) by cardinality matching; the within-district Mahalanobis caliper then retains 116 of these controls ($1722$ actas), so the estimation sample comprises 129 matched locations---13 treated and 116 control---with $1909$ actas in total. Table (ref) reports the standardized mean difference (SMD) for each matching covariate before and after matching. The largest absolute imbalance falls from $0.79$ in the full control pool to $0.100$ in the matched sample, and every covariate clears the imposed tolerance, so the treated and matched-control distributions are comparable on observables.
Table (ref) reports the two estimators of $\tau_c$ for the seven leading candidates: the acta-weighted difference-in-means estimator $\widehat\tau_c^{\,\mathrm{DM}}$ of (ref) and regression-adjusted estimator $\widehat\tau_c^{\,\mathrm{R}}$ of (ref) with cluster-robust inference. The two agree in sign and magnitude for every candidate, so the adjustment for residual demographic imbalance does little work and the global balance of (ref) carries the identification.
Three patterns emerge. First, the three candidates the flash estimates rendered viable---Nieto, L\'opez Aliaga, and S\'anchez---all gain significant vote share among Monday voters relative to matched Sunday locales, and the leading non-viable candidates---\'Alvarez, Belmont, and L\'opez Chau---all lose share. This is the sign pattern the strategic-reallocation logic of the model predicts (H(ref)--H(ref), H(ref)). Second, the incumbent runoff entrant Fujimori gains a small but significant share ($+2.6$ percentage points), consistent with the weak-reinforcement prediction of H(ref). Third, the reallocation is overwhelmingly concentrated on Nieto, whose $+17.3$-point gain is three to seven times the gain of any other candidate. The seven leading candidates' effects sum to $+0.134$, so the remaining $13.4$ points of displaced Monday share are drawn from the long tail of minor candidates, consistent with the adding-up implication of H(ref).
The estimated magnitudes track the affinity ranking of Section (ref). The signs reproduce the model's core division of the ballot---share moves toward the viable candidates and away from the rest---and the within-viable ordering is Nieto $>$ L\'opez Aliaga $>$ S\'anchez, exactly as H(ref)--H(ref) anticipate. Nieto, the less extreme option standing outside the discredited left--right blocs, combines the viability the flash conferred with the highest expressive match in the displaced electorate and captures the largest gain; L\'opez Aliaga, viable and ideologically aligned but weighed down by his identification with the congressional pact, gains an intermediate share; S\'anchez, viable nationally but expressively distant in this right-leaning metropolitan sample, gains the least. Because H1–H3 are ordinal, the size of the gap is not a departure from the predictions but a quantitative reading of them: Nieto's gain is several times López Aliaga's, which underscores how strongly the off-axis clean attribute operates in the displaced electorate relative to the two right-aligned alternatives. The aggregate magnitudes are consistent with the viability--affinity interaction but cannot, on their own, separate affinity from viability: a candidate's national viability and his local gain need not coincide, as the S\'anchez case makes clear.
The headline pattern survives every exercise of the sensitivity battery of Section (ref).
\paragraph{Location-weighted estimand.} Re-weighting so that each treated location counts equally, rather than in proportion to its number of actas, leaves every sign and significance unchanged (Table (ref)). The point estimates are close to the acta-weighted ones---Nieto $+0.176$, L\'opez Aliaga $+0.054$, S\'anchez $+0.025$---so the result is not driven by the larger treated locations.
\paragraph{Leave-one-treated-location-out.} Recomputing the estimator thirteen times, excluding one treated location at a time, produces no sign reversal and no loss of significance at the $5\%$ level for any candidate (Table (ref)). The ranges are tight: Nieto's effect stays within $[0.170,\,0.178]$, and the group-by-group decomposition is broad-based---each viable candidate's within-group contrast is positive in at least $11$ of the $13$ groups and no single group accounts for more than a quarter of any aggregate effect.
\paragraph{Balance tolerance.} Re-solving the cardinality-matching problem (ref) at $\delta \in \{0.05, 0.10, 0.15\}$ moves the estimates negligibly (Table (ref)); the result is not an artifact of a single tolerance. Tighter tolerances select fewer cardinality-selected cardinality-selected controls (163, 178, 200 locations) at a better post-match balance, and all estimates remain significant at the $1\%$ level.
\paragraph{Rosenbaum bounds.} Under the matched-group sensitivity framework described in Section (ref), Table (ref) reports the worst-case $p$-value associated with each electoral outcome as the sensitivity parameter $\Gamma$ increases. For each outcome, the procedure evaluates how large an unobserved source of selection would need to be to overturn the inference obtained from the matched-group treatment--control contrasts. The reported value $\Gamma^\star$ corresponds to the largest degree of hidden bias for which the sharp null hypothesis of no treatment effect continues to be rejected at the $5\%$ significance level. Larger values of $\Gamma^\star$ indicate greater robustness to unobserved confounding, as the estimated effect remains statistically significant even when matched voting locations are allowed to differ substantially in their odds of treatment exposure due to factors not captured by the observed covariates. Conversely, values of $\Gamma^\star$ close to one suggest that relatively small departures from the selection-on-observables assumption would be sufficient to eliminate statistical significance.
\textcolor{red}{Note: The Rosenbaum bounds reported in this section are preliminary. The aggregated signed-score statistic is appropriate for matched pairs but does not faithfully characterize sensitivity in $1{:}k$ sets with a varying number of controls: in a set with $p_i$ controls, permuting the treatment assignment yields $1+p_i$ distinct contrasts rather than a sign flip of the average contrast over $\pm\hat\Delta_{i,c}$. The current implementation also decouples the inference base (within-group permutation) from the sensitivity base and does not reproduce the permutation $p$-value at $\Gamma=1$. This will be corrected in the next version with a faithful matched-sets implementation: per-location scores, binary worst-case within-set tilt under $\Gamma$, and exact combination across the 13 independent groups.}
We study how election-night flash estimates reshape voting in a fragmented plurality election. A Bayesian-updating model in which voter $j$'s payoff from candidate $c$ is $\alpha_j M_j(c) + \beta_j\, M_j(c)\,\mathbf{1}\{\cdot\}$ yields a viability--affinity interaction: a public signal raises perceived viability $q_c$ uniformly, but the vote it reallocates is scaled by the expressive match $M_j(c)$, making reallocation a within-neighborhood phenomenon rather than a mechanical flow toward whoever the flash names. We exploit a natural experiment---the Jurado Nacional de Elecciones extended voting in 13 Lima locations to the Monday after election day, exposing 187 actas to the overnight Ipsos and Datum estimates while comparable Sunday locations were not---and estimate candidate-level effects on 2026 first-round shares by cardinality matching (129 matched locations: 13 treated and 116 control; 1909 actas) with cluster-robust inference. The acta-weighted difference-in-means and within-group regression estimators agree in sign and magnitude for every candidate, so the adjustment for residual demographic imbalance does little work. Identification, however, rests on within-group comparability rather than on the pooled balance alone: the estimators weight each group's control actas by $w_i/N_i^C$, which departs from the equal-acta weighting $1/N_{\mathcal{C}^\star}$ of the pooled SMDs in Table (ref) once group sizes differ. Balance under this weighting is secured by the within-district Mahalanobis caliper of Section (ref), which keeps each treated location close to its own matched controls, while (ref) delivers the global marginal balance.
Our contribution is to identify the causal effect of a late viability signal on vote reallocation and to show that it runs through the local distribution of expressive affinities, not the signal alone; the model is the organizing device for this claim, not a stand-alone result.
Information has a significant, robust effect. The flash raised the shares of the three candidates it rendered viable (Nieto, L\'opez Aliaga, S\'anchez; H(ref)--H(ref)), weakly reinforced Fujimori ($+2.6$ points; H(ref)), and lowered the non-viable candidates' shares (H(ref)); every effect is significant at $1\%$ and survives the leave-one-out, weighting, tolerance, and Rosenbaum checks. The magnitudes follow $M_j(c)$: because $q_c$ moves uniformly while the reallocation is scaled by match, the ordering Nieto $>$ L\'opez Aliaga $>$ S\'anchez tracks affinity, not national viability. Nieto, the off-axis “clean” option with the highest match, takes the bulk ($+17.3$ points); L\'opez Aliaga's match is depressed by his tie to the congressional pact ($+5.1$); S\'anchez, the eventual runoff entrant, is expressively most distant in this right-leaning sample and gains least ($+2.7$). A candidate can thus advance nationally on a signal yet move few local votes, because viability travels with the signal while its reallocation is governed by local affinity.
\printbibliography