Don't Judge a Policy by its Politician

All participants read twenty policy proposals quoted verbatim from the official government plans that the two leading presidential candidates registered with Brazil's electoral court (TSE), and rate their agreement with each on a five-point scale.

Three interventions are assigned at random and independently of one another.

Candidate attribution. Half of participants see each proposal preceded by the name of the candidate who proposed it. The other half see the identical proposal text with no name attached. The wording of the proposals, their order, and the response scale are the same in both versions; only the candidate's name is added or omitted.

Depolarization video. Half of participants watch a short educational video (about four minutes) on the social costs of political polarization before rating any proposals. The other half proceed directly to the proposals.

Alignment feedback. After rating all twenty proposals, half of participants are shown two percentages: the share of each candidate's proposals they agreed with, presented as their agreement with the platform of the left (Lula) and the platform of the right (Flávio Bolsonaro). The other half are shown no result and proceed directly to the closing questions. Both percentages are computed only from the participant's own ratings and are accurate. Participants are informed in the consent form that this result should not be interpreted as a voting recommendation; no candidate is recommended to any participant at any point.

Immediately before any result is displayed, all participants state what they believe those two percentages will be.

2026-09-25

2026-11-20

Primary outcomes, all measured in Wave 1.

Item-level agreement. The participant's rating of each of the twenty proposals on a five-point scale, recoded so that higher values indicate greater agreement (4 = strongly agree through 0 = strongly disagree). Twenty observations per participant.

Alignment rate. The participant's computed agreement with each candidate's platform, on a 0 to 100 scale, constructed from the ten items belonging to that candidate. Two values per participant, together with the difference between the rate for the candidate they preferred at baseline and the rate for the other candidate.

The confirmatory quantity is the effect of candidate attribution on these measures, conditional on who wrote the proposal: whether naming the author raises agreement with proposals written by the participant's preferred candidate relative to proposals written by the opposing candidate. Both endpoints are recorded for every Wave 1 completer and are unaffected by follow-up attrition.

Alignment rate. For participant $i$ and candidate $k$, responses to $k$'s ten proposals are scored 4 (strongly agree), 3 (agree), 2 (neutral), 1 (disagree), 0 (strongly disagree), summed, divided by 40, and multiplied by 100. The instrument records responses in the reverse order, with 1 for strongly agree, so the scale is inverted before scoring. This is the figure displayed to participants in the feedback arm. Where a participant has fewer than eight valid responses among a candidate's ten items, the alignment rate for that candidate is treated as missing in the analysis and is not imputed.

Authorship. An indicator equal to 1 if a proposal was written by the candidate the participant preferred at baseline, and 0 if written by the other candidate. Preference is taken from the baseline 100-point vote allocation: the preferred candidate is whichever of the two received strictly more points. Participants who divided their points equally between the two, or who placed their largest share on another candidate, on blank or null, or on abstention, have no preferred candidate; they are excluded from the primary specification, and we report the share excluded. As a prespecified alternative definition we map the five-category political self-placement to sides, assigning Lulista and non-Lulista left to Lula and Bolsonarista and non-Bolsonarista right to Flávio Bolsonaro, with independents excluded.

The primary endpoint combines these two: the effect of candidate attribution on agreement, estimated as the interaction between the attribution indicator and the authorship indicator. Authorship is constructed entirely from measures taken before randomization, so conditioning on it and interacting with it is legitimate.

Constructions used in secondary analyses

Misperception gap. The signed difference between the computed alignment rate and the belief the participant stated immediately before the feedback screen, for each candidate. Positive values mean the participant agreed with that candidate more often than they expected. The gap provides the identifying variation for the feedback analyses, since participants whose stated belief was already accurate learn nothing from the screen and the treatment is not expected to produce a constant shift. Unlike the authorship indicator, the gap is constructed from ratings collected after the attribution and depolarization treatments have been delivered, so it is post-treatment with respect to both, and specifications involving it are estimated within cells defined by those two factors.

Belief revision. The change in stated alignment belief between Wave 1, where it is elicited immediately before the feedback screen, and Wave 3, roughly seven weeks later. No belief is elicited after the feedback screen within Wave 1, so this measures persistence rather than immediate updating and bounds from below whatever revision occurred and subsequently decayed.

Vote-intention share. The points allocated to a given candidate in the baseline 100-point allocation. This is measured once, before randomization, and serves as the pre-treatment covariate used to define authorship and to adjust the Wave 2 vote analysis. Because the allocation sums to 100 by construction, the five components are not treated as independent quantities.

Wave 2, immediately after the election.

1. Self-reported vote, in five categories: Lula, Flávio Bolsonaro, another candidate, blank or null, did not vote.
2. Conviction in that vote decision, on a scale from 0 to 10.
3. Feeling thermometers toward the political left and toward the political right, each from 0 to 100.

Wave 3, roughly three weeks after the election.
4. Item-level agreement with the same twenty proposals, presented without candidate names to every participant.
5. Alignment rate with each platform, recomputed from those ratings.
6. Stated belief about own alignment with each platform, on two scales from 0 to 100.

Waves 1 and 3, closing battery.
7. Perceived purpose of the study, selected from six options.
8. Whether the participant left a free-text comment, and whether that comment names a candidate.
9. Two ratings of the survey itself, for quality and for neutrality, each from 1 to 5.

Vote. The primary coding is an indicator that the participant reports voting for the candidate who received the larger share of their baseline 100-point allocation. The full five-category variable is also reported, keeping blank or null and abstention as distinct categories rather than combining them.

Affective polarization. Constructed from the Wave 2 thermometers as the difference between the rating the participant gives to their own side and the rating they give to the other side. The thermometers refer to the political left and the political right rather than to the candidates or to their supporters. Sides are assigned from the baseline vote allocation, mapping a preference for Lula to the left and a preference for Flávio Bolsonaro to the right; participants with no preferred candidate at baseline are assigned using the five-category political self-placement, and independents are excluded.

The instrument contains no affective measure before treatment and none at the level of individual candidates or their supporters. There is therefore no manipulation check for the depolarization video, and this outcome is estimated as a between-arm difference in Wave 2 without baseline adjustment. A null result will not distinguish a video that failed to reduce animus from animus that does not move these behaviors, and we will report it on that basis.

Long-run policy preferences. Wave 3 item-level agreement, adjusted for the participant's Wave 1 rating of the same item, with item fixed effects and standard errors clustered by participant. Because Wave 3 presents the twenty proposals without labels to every participant, this measures underlying agreement with the proposals rather than labeled endorsement. For participants originally assigned to the identified arm, the change from Wave 1 to Wave 3 combines any persistence of treatment with the removal of the label, so that comparison is reported separately and descriptively.

Long-run beliefs. The change in stated alignment belief between Wave 1, where it is elicited immediately before the feedback screen, and Wave 3, roughly seven weeks later. No belief is elicited after the feedback screen within Wave 1, so this measures persistence rather than immediate updating and bounds from below whatever revision occurred and subsequently decayed.

Experimenter-demand diagnostics. Following Bursztyn, Haaland, Rao and Roth (2020), three endline measures are compared across arms: whether the participant left a free-text comment, whether that comment names a candidate, coded by a prespecified case-insensitive string match on both candidates' names and common variants, and the two survey quality and neutrality ratings. Balance across arms is evidence against differential inference about the study's intent.

Perceived purpose. The six response options include an accurate description of the study's aim. Because this item is administered in Wave 1, before two further contacts, some participants are shown a plausible statement of the hypothesis and are then recontacted. We test whether selecting that option in Wave 1 predicts the Wave 2 and Wave 3 outcomes, and whether the feedback effect differs between participants who selected it and those who did not.

The study is a single online survey session fielded in the weeks before the first round of Brazil's October 2026 presidential election, followed by two shorter contacts after it.

Invitations are sent by a partner online lottery platform to its user base, using the platform's own systems. The research team does not receive names, e-mail addresses, or account information; responses are linked across contacts by a pseudonymous code held by the platform.

Participants who consent, report a year of birth indicating they are 18 or over, and confirm they reside in Brazil answer questions on demographics and on their political self-placement and vote intention. Vote intention is recorded as an allocation of 100 points across the two leading candidates, another candidate, blank or null, and abstention. Two attention checks are administered at this stage, before any randomization.

Three treatments are then assigned at random and independently of one another: the depolarization video, candidate attribution on the proposals, and alignment feedback. This produces a 2 × 2 × 2 design with eight cells. Because the assignments are independent, each treatment can be evaluated on its own and in interaction with the others.

Participants then rate the twenty proposals, presented in an order randomized within participant. They state their expectations about their own alignment with each platform, receive the computed percentages or not according to assignment, and close the session with questions about the survey itself. No further measures are collected in this session.

A short follow-up shortly after the election records the reported vote, conviction in that decision, and feeling toward the political left and the political right. A final contact about three weeks later asks participants to rate the same twenty proposals again, this time with no candidate names shown to anyone, and to restate their expected alignment. A full debriefing is then sent to every participant who consented, including those who did not complete the follow-ups.

Not available

Randomization is performed automatically by the survey software (Qualtrics) at the moment each respondent reaches the randomization point, after the baseline module and before any treatment is delivered. Three separate randomizers operate independently of one another, each assigning one of two conditions with equal probability using Qualtrics' "evenly present elements" allocation, which keeps the counts in the two conditions close as responses accrue rather than drawing each assignment independently. Assignment is therefore independent across the three factors within a respondent, and balanced but not strictly independent across respondents; inference treats the allocation as simple randomization, which is conservative for this design.

No stratification on respondent characteristics is used, and no blocking on region, demographics, or political self-placement. Each assignment is written to the response as an embedded data field, giving three binary indicators that are stored with every record and used to define the analysis arms.

Respondents reach the survey through an individual link carrying a pseudonymous code supplied by the partner platform. Where more than one response is recorded against the same code, the first response that passes the pre-randomization screening criteria is retained and later responses are discarded; the number of such cases is reported.

Individual respondent. All three treatments are randomized at the individual level, and there is no group- or cluster-level assignment.

A second randomization operates within respondent rather than across respondents: the order in which the twenty proposals are presented is randomized for each participant. This is not a treatment, and the realized order is stored so that position within the sequence can be used as a covariate and tested for order effects.

Note that the unit of observation for the primary outcome is the proposal rather than the respondent, giving twenty observations per assigned unit. Standard errors in the item-level models are therefore clustered by respondent, which is the level at which treatment is assigned.

No

Not clustered; randomization is at the individual level, so the number of clusters equals the number of individuals.

2,500 to 5,000 individual respondents expected to be randomized and analyzed, drawn from approximately 25,000 platform users invited. Randomization occurs after consent, eligibility screening, and two attention checks, so the invited figure is an upper bound on the number of units entering the experiment rather than the planned number itself.

2,500 to 5,000 individual respondents, the same as the number of clusters since the design is not clustered, drawn from approximately 25,000 platform users invited. The primary outcome is measured at the level of the proposal rather than the respondent, with twenty proposals rated per person, giving approximately 50,000 to 100,000 proposal-level observations in Wave 1. Follow-up contacts are smaller. Assuming 50 percent retention, Wave 2 yields roughly 1,250 to 2,500 respondent-level observations on the reported vote. Assuming 40 percent retention, Wave 3 yields roughly 1,000 to 2,000 respondents, each rating the same twenty proposals again, for approximately 20,000 to 40,000 proposal-level observations.

Equal allocation across a 2 × 2 × 2 design, one eighth per cell. There is no pilot; the study is fielded once.

Expected sample: 2,500 to 5,000 participants, drawn from approximately 25,000 invited platform users.

Per factor, each arm receives half the sample, giving 1,250 to 2,500 participants each: depolarization video / no video; candidate-identified proposals / blind proposals; alignment feedback shown / not shown.

Per cell, one eighth of the sample, giving approximately 310 to 625 participants each. The eight cells are: no depolarization–blind–no feedback; no depolarization–blind–feedback; no depolarization–identified–no feedback; no depolarization–identified–feedback; depolarization–blind–no feedback; depolarization–blind–feedback; depolarization–identified–no feedback; depolarization–identified–feedback.

All calculations assume 80 percent power, a 5 percent two-sided test, and equal allocation, using MDE = 2.80 × σ × √(2/n) per arm. Item-level models cluster standard errors by respondent, so the effective unit is the participant, with twenty proposals each. Assumed standard deviations are stated so they can be checked against realized data. Retention is assumed at 50 percent for Wave 2 and 40 percent for Wave 3. The range in each row spans the expected sample of 2,500 to 5,000 completions. Primary outcome, attribution effect on alignment. Estimated on the full Wave 1 sample with depolarization fixed effects, giving 1,250 to 2,500 participants per attribution arm. Alignment rate, per candidate, 0–100 scale, assumed σ = 18 points: minimum detectable effect of 2.0 to 1.4 points, or 0.11 to 0.08 standard deviations. Difference between own-candidate and opposing-candidate alignment rate, assumed σ = 25 points: 2.8 to 2.0 points, or 0.11 to 0.08 standard deviations. Item-level agreement, attribution by authorship interaction, five-point scale. The relevant quantity is the within-participant difference between mean agreement with own-candidate and opposing-candidate proposals, assumed σ = 1.1 scale points given an item-level σ of 1.3 and moderate within-person correlation: 0.12 to 0.09 scale points, or 0.09 to 0.07 item-level standard deviations. Secondary outcomes. Feedback effect on the reported vote, Wave 2, binary indicator with p ≈ 0.5 and σ = 0.5, 625 to 1,250 per arm: 7.9 to 5.6 percentage points. Feedback effect on Wave 3 policy preferences, item level, adjusted for the Wave 1 rating of the same proposal, assumed residual σ = 0.54 at the participant level, 500 to 1,000 per arm: 0.10 to 0.07 scale points, or 0.07 to 0.05 item-level standard deviations. Feedback effect on long-run alignment beliefs, Wave 1 to Wave 3, 0–100 scale, assumed σ = 20 points: 3.5 to 2.5 points. Depolarization effect on affective polarization, measured as the difference between the Wave 2 left and right thermometers, assumed σ = 40 points: 6.3 to 4.5 points. Exploratory. The interaction between feedback and the misperception gap requires roughly four times the sample of a main effect of equal standardized size. On the Wave 2 vote this gives a minimum detectable differential of approximately 16 to 11 percentage points per standard deviation of the gap. This is larger than any effect the information-provision literature reports, so the interaction is prespecified as exploratory rather than confirmatory. Power for it depends on the dispersion of the misperception gap rather than on sample size alone: if most participants prove approximately correct about their own alignment, the interaction is weakly identified at any sample size. The distribution of the gap is the first quantity we report.