Economic Reform and Electoral Accountability: Evidence from Argentina

We study whether and how exposure to a clear incumbent-driven program of economic adjustment shapes electoral reward and punishment, and how attribution, political identity, a change in aggregate information, and individual characteristics moderate that process. To identify these mechanisms the design combines three elements. First, we are interested in understanding how the economy shapes voting (economic voting). For this we estimate a household-level change in welfare (Δ) that proxies each respondent’s exposure to the incumbent’s economic program. Second, as economic exposure and political affinity are hard to isolate, we rely on randomized economic and political primes using a randomized within-survey attribution/identity prime with four arms. Finally, we exploit a natural information shock generated by a provincial legislative contest together with geographical variation in exposure.

Estimating household exposure (Δ). We define the household-level treatment as a latent Δ that captures the change in household economic welfare plausibly attributable to the incumbent’s program (e.g., changes in consumption, real income). To retrieve Δ we estimate group-specific pre/post welfare changes using first-period national household survey data and a second period consisting of our VOCES interviews, following the approach in Del Canto, Grigsby, Qian, and Walsh (2025). Groups are formed on the basis of observable covariates. For each group we compute the average pre–post welfare change and then match the resulting group-level Δ to VOCES respondents with the same observable profile. Conceptually this treats the macro policy as a common shock whose realized incidence varies idiosyncratically across households; empirically it is analogous to a shift-share construction and yields plausibly exogenous cross-sectional variation in exposure conditional on observables. We present robustness checks that vary group definitions and imputation methods to assess sensitivity. We will validate this measure in two ways: first, by using an open-ended question that asks respondents how the decline in inflation has affected their lives; and second, by comparing it with survey variables that capture changes over time in income and other household characteristics.

Randomized primes (within-survey intervention). Within VOCES respondents are randomly assigned to one of four arms: (i) an economic attribution prime that prompts respondents to link the observed drop in inflation to the incumbent’s program (a brief, neutral reminder emphasizing causality: policy → observed macro change); (ii) a political-identity prime that asks respondents to recall or reflect on long-standing partisan attachments (framed, for example, as a short reminder about family political history); (iii) a combined prime receiving both the economic and identity reminders; and (iv) a control arm receiving no reminder. Random assignment is implemented orthogonally to respondent observables, interviewer, interview location, and interview timing, producing exogenous variation in attribution and identity salience. The design allows estimation of main effects of each reminder and of interactions between reminders and the imputed Δ. Inference proceeds under an intent-to-treat (ITT) framework.

To avoid post-treatment bias when estimating heterogeneous effects, we also measure long-standing partisan identity using pre-treatment proxies (historical parental party affiliation). We treat these parental-affiliation measures as stable, pre-existing markers of political identity that are not plausibly affected by the policy shock; they therefore serve as moderators rather than outcomes.

Interacting Δ and randomized primes allows us to test, for example, whether durable partisan attachments (e.g., strong Peronist/Kirchnerist identity) attenuate or amplify the translation of material gains and losses into electoral reward or punishment.

Natural information shock and geographical variation. We exploit the Buenos Aires provincial legislative contest (7 September 2025) as a quasi-experimental information shock that altered the informational environment about the incumbent’s political viability to explore how individuals update their perceptions of the economy and adjust their political preferences accordingly. Individuals generally have accurate information about their personal economic conditions but face uncertainty about the broader economic context. In this setting, voters may reward the incumbent only if they believe the national economy is performing well. The recent economic developments in Buenos Aires provide new information that helps respondents update their beliefs about the overall economy. If people observe that others are struggling, they may infer that the economy is not improving and therefore become less likely to support President Milei.

Respondents interviewed before and after the contest faced different public signals about electoral support for the incumbent; we use interview timing to capture belief updating around the event. In addition, because the contest was geographically localized, we use respondents’ interview location (Buenos Aires vs. other provinces) as a second layer of variation. Combining temporal and spatial variation with the randomized primes and the imputed Δ allows us to assess how public signals, individual exposure, and attribution/identity interventions jointly determine political responses.

Conceptually the design is a mixed strategy that brings together observationally heterogeneity in welfare change, randomized within-survey treatments (primes) and quasi-experimental temporal variation (pre/post contest). The primary strategy estimate (i) the association between economic exposure and political outcomes (vote intention, approval, willingness to reward/punish), (ii) the causal effect of the randomized primes, and (iii) the causal effect of the temporal shock.

2025-08-01

2025-09-13

Our main outcomes capture behavioral and attitudinal responses to the incumbent’s economic program.

1. Overall Assessment

a. Vote intention:
Respondents report which political party they plan to vote for in the upcoming elections (open, precoded list). We recode this variable into an indicator for incumbent vote intention (La Libertad Avanza = 1, all others = 0). Thinking about the upcoming legislative elections in October, which political party do you plan to vote for in the election for the Chamber of Deputies?
b. Overall evaluation/rating:
Using a five-point scale (1 = very bad to 5 = very good), we build both the raw ordinal measure and a standardized continuous version (0–1) reflecting overall program approval. Overall, how would you rate the performance of the current president, Javier Milei?

2. Political Outcomes

a. Affective polarization. Respondents express their feelings toward voters of Milei and Massa on a 0–10 scale (0 = very negative, 10 = very positive). We compute an affective-polarization index defined as the difference between out-group and in-group ratings.
b. Assessment of Milei in non-economic dimensions:
i. Liberal democracy. Respondents rate the current situation of the country in three areas: (i) protection of women’s and minority rights, (ii) freedom of expression, and (iii) human rights. Each is measured on a five-point scale from very bad to very good; we construct a standardized index averaging the three items.
ii. Security. Crime and violence have increased in the last twelve months? and others
iii. Overall evaluation of democracy. How satisfied are you with the way democracy works in Argentina? and others

3. Economic Outcomes

a. Now, thinking about the Argentine population in general, and using a scale from 1 to 5, where 1 means very negatively and 5 means very positively, how do you think disinflation has affected the population?
b. Thinking about the 2025 legislative elections, how would you say the decline in inflation has influenced your voting decision?
c. How would you rate the current government’s economic program?

We will also measure institutional trust and perceived corruption as secondary outcomes.

We will construct a standardized index averaging the items included in each of those modules.

1. The Sample

The sample was stratified to ensure representativeness of socioeconomic levels within the urban population. Survey respondents are adults between the ages of 18 and 50, randomly selected from the household members in this age range. The survey is designed to take approximately 35 minutes to complete. The survey yielded around 2,622 effective observations.

We took multiple measures to increase response rate:
- If the initially selected individual is unavailable, another individual from the same household in the same age range may be selected using the same randomization criteria. A maximum of two replacements is allowed. If the third randomly selected individual does not respond, the household will no longer be part of the sample.
- An individual is considered unavailable under the following circumstances: refusal to answer, absence (for in-person surveys), failure to answer the phone (in phone surveys), or health or disability conditions that prevent participation. Only one individual per household will be surveyed.
- A protocol will be implemented to minimize non-response rates and ensure quality data collection.

2. Fieldwork
The firm received the initial version of the questionnaire and adjusted it for local language usage. This is followed by a pilot survey administered to a small number of households. This ensures that the survey concepts are well understood across different contexts and socioeconomic backgrounds, and helps identify issues with survey flow and respondent comprehension.

Based on the findings of the pilot survey, the questionnaire was refined, and the official survey is launched. Fieldwork did not last no longer than four weeks.

3. Current Progress
The survey was piloted on July 10–11, 2025. Fieldwork began on August 1 and was completed on September 13, 2025. At the time of writing, data collection had been completed. This PAP is being registered prior to any data analysis.

4. Analysis

Our first specification estimates how economic exposure shapes voting behavior:

Y_i = \beta_0 + \beta_1 , \text{EconomicExposure}_i
+ \gamma_1 , \text{PoliticalAffinity}_i
+ \lambda , (\text{EconomicExposure}_i \times \text{PoliticalAffinity}*i)
+ \alpha_j + \delta X_i' + \varepsilon*{i,p}

where:

( Y_i ) is the outcome variable.
( \alpha_j ) captures questionnaire version fixed effects.
( \varepsilon_{i,p} ) is the error term.
( \text{EconomicExposure}_i ) measures respondent i’s exposure to the incumbent’s economic program.
( \text{PoliticalAffinity}_i ) captures attachments to the incumbent or challenger (in this case, Peronismo).
( X_i' ) is a vector of sociodemographic controls.

We will estimate coefficients using OLS on the full sample.

Our second specification isolates the effects of economic and political exposure using the randomization exercise:

Y_i = \beta_0
+ \beta_1 , \text{BoosterEconomic}_i
+ \beta_2 , \text{BoosterPolitical}_i
+ \beta_3 , \text{BoosterBoth}*i
+ \delta X_i' + \varepsilon*{i,p}

where:

( \text{BoosterEconomic}_i ) is a binary indicator equal to 1 if the respondent was assigned to the economic exposure treatment.
( \text{BoosterPolitical}_i ) and ( \text{BoosterBoth}_i ) indicate assignment to the political exposure and combined treatment arms, respectively.

Our main goal is to examine how exposure to an economic program affects electoral behavior. The central trade-off lies between economic performance (how individuals perceive their economic situation) and political affinity (their partisan attachments). The experimental treatments act as reminders or priming interventions designed to make these dimensions more salient to respondents, thereby amplifying the weight of economic or political considerations in their decision-making. We expect heterogeneous effects, since not all respondents will respond equally to these primes. Our analysis will therefore explore how treatment effects vary with baseline economic exposure and political affinity, allowing us to trace how economic information and partisanship jointly shape political behavior.

(a) Varying Political Affinity Levels

Y_i &= \beta_0
+ \beta_1 , \text{BoosterEconomic}_i
+ \beta_2 , \text{BoosterPolitical}_i
+ \beta_3 , \text{BoosterBoth}*i
+ \gamma_1 , \text{PoliticalAffinity}_i
+ \gamma_2 , (\text{PoliticalAffinity}_i \times \text{BoosterEconomic}_i)
+ \gamma_3 , (\text{PoliticalAffinity}_i \times \text{BoosterPolitical}_i)
+ \gamma_4 , (\text{PoliticalAffinity}_i \times \text{BoosterBoth}_i)
+ \delta X_i' + \varepsilon*{i,p} \

(b) Varying Economic Exposure Levels

Y_i &= \beta_0
+ \beta_1 , \text{BoosterEconomic}_i
+ \beta_2 , \text{BoosterPolitical}_i
+ \beta_3 , \text{BoosterBoth}*i
+ \gamma_1 , \text{EconomicExposure}_i
+ \gamma_2 , (\text{EconomicExposure}_i \times \text{BoosterEconomic}_i)
+ \gamma_3 , (\text{EconomicExposure}_i \times \text{BoosterPolitical}_i)
+ \gamma_4 , (\text{EconomicExposure}_i \times \text{BoosterBoth}_i)
+ \delta X_i' + \varepsilon*{i,p} \

Similarly, our third specification isolates the effects of economic and political exposure using the quasi-exogenous variation from the elections:

Y_i = \beta_0
+ \beta_1 , \text{Elections_post}_i
+ \delta X_i' + \varepsilon*{i,p}

where:

( \text{Elections_post}_i ) is a binary indicator equal to 1 if the respondent was surveyed after the elections.

We then combine both approaches to test heterogeneous effects:

(a) Varying Political Affinity Levels

Y_i &= \beta_0
+ \beta_1 , \text{Elections_post}_i
+ \gamma_1 , \text{PoliticalAffinity}_i
+ \gamma_2 , (\text{PoliticalAffinity}_i \times \text{Elections_post}_i)
+ \delta X_i' + \varepsilon*{i,p} \

(b) Varying Economic Exposure Levels

Y_i &= \beta_0
+ \beta_1 , \text{Elections_post}_i
+ \gamma_1 , \text{EconomicExposure}_i
+ \gamma_2 , (\text{EconomicExposure}_i \times \text{Elections_post}_i)
+ \delta X_i' + \varepsilon*{i,p} \

For robustness, we will estimate three parallel specifications:

1. A model without controls.
2. A model including only the minimal set of controls needed after assessing sample balance.
3. A model using double post-LASSO–selected covariates.

Since treatment is assigned at the individual level and the design does not involve multiple time periods, clustering is unnecessary. We will use Eicker–Huber–White robust standard errors, as discussed in \cite{abadie2023clustering}.

Given the presence of multiple outcome variables, we will also adjust for multiple hypothesis testing following standard corrections.

Heterogeneity by various indexes:

1) Institutional Trust

How much do you trust the National Government? (1 = Trust nothing — 5 = Trust a lot).
How much do you trust the Local / provincial government? (1–5).
How much do you trust the Congress? (1–5).
How much do you trust the Judiciary (judges and courts)? (1–5).
How much do you trust political parties? (1–5).
How much do you trust the police? (1–5).
How much do you trust the electoral process? (1–5).

2) Pro-sociality (Redistribution preferences + Perceptions of inequality)

On a ladder from 1 (poorest) to 10 (richest), where would you place yourself? → (item for relative position)
Up to which step on the ladder should government assistance reach? (write a number 0–10)
From which step on the ladder should families start paying taxes? (write a number 0–10)
To what extent do you agree that the level of inequality in Argentina is acceptable? (1 = completely unacceptable — 5 = completely acceptable)
Do you think a person's wealth is due mainly to hard work or inherited advantages? (options: “mostly hard work” / “mostly advantages” / “both equally”)

3) Ideology

On a scale from 0 (Left) to 10 (Right), where would you place your political views? (0–10)

4) Commitment to Liberal Democracy

How essential are each of the following to democracy? (1 = not essential, 5 = essential) — list items individually:
- Freedom to criticize the government.
- Periodic free and fair elections.
- A Congress that checks presidential power.
- Political parties.
To what extent do you agree with: “Democracy may have problems, but it is better than any other form of government.” (1–5)
How satisfied are you with the way democracy works in Argentina? (1 = very dissatisfied — 4 = very satisfied)

5) Main National Problem (binary: economy = 1)

“In your opinion, what is the most important problem facing the country today?”

6) Anti-establishment Sentiments

To what extent do you agree:
“The main division in society is between citizens and the economic and political elites.” (1–5)
“The country’s economy is rigged to favor the rich and powerful.” (1–5)
“To fix the country, we need a strong leader willing to break the rules.” (1–5)

7) Exposure to Public Services

How would you rate the quality of primary and secondary education provided by the state? (1 = very bad — 5 = excellent)
How would you rate health services (vaccinations, healthcare) in your area? (1–5)
How would you rate urban public transport? (1–5)
Have you paid for the following services privately in your household? (Yes / No): primary education, higher education, healthcare. (binary)

Complete randomization done in office by a computer

Individual (survey respondent)

No

NA

2,622 respondents

Our sample is distributed as follows:
Economic moderator: 691
Political-affiliation moderator: 604
Both: 675
Control: 652

We calculate the Minimum Detectable Effect (MDE) for each type of outcome variable. The MDE represents the smallest true effect size that can be detected with a specified level of power, given the study's sample size and design. We opted for this approach because we have already established a fixed sample size of N=2,622 in each country -due to budget restrictions-, rather than determining the sample size that maximizes statistical power. Our analysis encompasses various outcome variables, including binary variables, ordinal variables (scale: 1 to 5, 1 to 3, -10 to 10) and continuous variables (0 to 1). To accommodate this diverse set of outcomes, we calculated the MDE using each of the 5 types of outcomes. Furthermore, we simulated various potential distributions for these four types of variables, exploring different combinations of means and standard deviations. This comprehensive approach allows us to assess the study's sensitivity to detect effects across a range of plausible scenarios. Notably, our calculations indicate that our study design has sufficient statistical power to detect even relatively small effect sizes. Our MDEs resulting from the most conservative assumptions are: for binary outcomes is 0.04, for ordinal variables (1 to 5) is 0.12, for ordinal variables (1 to 3) is 0.06, , for ordinal variables (-10 to 10) is 0.32 and for continuous variables is 0.02.