Evidence Provision, Defendant Race, and Criminal Justice Punitiveness: A Survey Experiment in Brazil

Participants are randomly assigned to one of two conditions. The treatment group receives a summary of a peer-reviewed study (Dobbie, Goldin, and Yang, 2018, American Economic Review) on the effects of pretrial detention on recidivism and labor market outcomes. The control group completes the same survey without receiving any research summary. To ensure comparability across arms, Bilendi applies identical non-interlocking demographic quotas on gender, age, region, race, and income independently to each survey link, with target proportions specified separately for each arm.

2026-03-20

2026-03-27

Bail recommendation for the theft vignette (pretrial release vs. pretrial detention) — binary
Bail recommendation for the knife assault vignette — binary
Sentencing recommendation for the theft vignette — binary (most punitive (lenient) option vs. any less punitive option) and ordinal (1=most lenient, 3=most punitive) as robustness check
Sentencing recommendation for the knife assault vignette — binary and ordinal
Prior and posterior recidivism beliefs: estimated number of reoffenders out of 10 defendants released before trial (0–10 slider), elicited before the intervention for all participants and again after the intervention for treatment participants who report updating their beliefs
Heterogeneous treatment effects by defendant race (Black vs. White images shown)

Bail recommendation is binary (1 = pretrial detention, 0 = release). Sentencing recommendation is analyzed primarily as a binary variable equal to 1 if the participant selected the most punitive sentencing option and 0 otherwise. As a robustness check I estimate an ordered probit treating the full ordinal scale (1=most lenient to 3=most punitive) as the outcome.
The theft vignette is the primary outcome measure since the evidence directly concerns non-violent property crimes of the type represented in that vignette. The knife assault vignette serves as a placebo/spillover test: if treatment effects are content-specific, the effect should be smaller for the knife assault, which involves a violent crime not covered by the evidence.
The recidivism belief is elicited before the intervention for all participants as a prior. For treatment participants who report updating their beliefs (yes to the uptake question), it is elicited again as a posterior. The treatment effect on beliefs is the posterior minus the prior within the updating subsample of the treatment group. For treatment participants who report not updating and for control group participants, the posterior is assumed equal to the prior.
Heterogeneity analyses interact the treatment indicator with defendant race (Black=1).

Support for CRV petition (nationwide expansion of prison occupancy regulation) — binary (yes/no/unsure)
Budget allocation to prison expansion vs. rehabilitation programs — continuous (0–11 BRL billions slider)
Perceived likelihood that the theft vignette defendant has a prior criminal record — 5-point Likert scale
Perceived likelihood that the theft vignette defendant will commit a violent crime in the future — 5-point Likert scale
Confidence in prior (and posterior if applicable) recidivism beliefs — 5-point Likert scale, elicited immediately after the prior/posterior belief slider
Perceived adequacy of the evidence study methodology — 5-point scale, elicited retrospectively at the end of the survey for treatment participants only
Estimated reduction in reoffending for Black defendants specifically - 5 point scale (ranging from increase in recidivism to a reduction larger than the overall reduction in the evidence)
Estimated reduction in reoffending for White defendants specifically
Heterogeneous treatment effects by respondent race, other respondent characteristics, and evidence uptake (treatment group only)

CRV petition and budget allocation measure downstream policy preferences beyond the immediate vignette decisions. The budget allocation variable is recorded as BRL billions allocated to prison expansion (0–11), with the rehabilitation allocation as the arithmetic complement. Both capture whether treatment shifts broader punitive vs. rehabilitative orientations beyond the specific vignette context.
Mechanism vignette outcomes (perceived prior criminal record and perceived future violent crime) are always elicited using the theft vignette defendant image. These capture the racial attribution mechanism: whether defendant race affects criminality attributions independently of the crime described.
Confidence in prior recidivism belief serves a dual role. First, it enters the Bayesian updating model as a measure of prior precision: participants more confident in their prior should update less in response to the same evidence signal. Second, it serves as a balance check on prior belief strength across treatment and control groups, since it is elicited before treatment assignment affects outcomes.
Retrospective evidence quality rating is collected at the very end of the survey, after all outcome measures, for treatment participants only. Because it is elicited post-treatment and post-outcomes, it cannot serve as a pre-outcome signal precision parameter. Instead it is used as: (a) a moderator of treatment effects — participants who retrospectively rate the evidence as higher quality should show larger treatment effects; (b) an input into a correlational Bayesian decomposition alongside prior confidence; and (c) a measure of evidence credibility heterogeneity across respondent subgroups.
Evidence applicability by defendant race (estimated reduction for Black defendants and for White defendants separately) is elicited for treatment participants only, after the vignettes. These capture whether participants perceive the US evidence as differentially applicable depending on defendant race — a direct test of racialized evidence processing. The difference between estimated reductions for Black vs. White defendants is the key variable of interest here.
Heterogeneity analyses interact the treatment indicator with respondent race (collapsed to Black/Pardo vs. White) and/or other respondent characteristics. Evidence uptake (binary, treatment group only) is additionally used to split the treatment group into updaters and non-updaters for exploratory subgroup analysis.

Two-arm randomized survey experiment with a nationally representative quota sample of Brazilian adults (N=550), recruited via the Bilendi online panel. Participants are randomly assigned to treatment (N=275) or control (N=275), with with independant demographic quota balancing applied to each treatment arm. All participants complete demographic questions, a prior recidivism belief elicitation with a confidence rating, an attention check, and two case vignettes (theft followed by knife assault) requiring bail and sentencing recommendations, followed by a mechanism vignette and stated policy preference questions. The treatment group additionally receives a summary of peer-reviewed evidence on pretrial detention effects between the prior belief elicitation and the vignettes. Defendant race (all-Black or all-White AI-generated images) is independently randomized at the individual level.

Not available

Individual-level randomization implemented via JavaScript executed within the Qualtrics survey platform at survey entry. Treatment vs. control assignment is determined by two separate survey links distributed by the panel company Bilendi, with independant demographic quota balancing applied to each link. Within-survey randomization (defendant race) uses a pseudorandom number generator seeded at runtime.

Individual. Each participant is independently randomized to treatment or control via survey link assignment, and to defendant race condition via in-survey JavaScript randomization.

No

550 individuals (no clustering; unit of randomization equals unit of observation)

550 individuals; 1,100 vignette-level observations (2 vignettes per participant)

275 individuals in treatment arm; 275 individuals in control arm

For the primary binary outcome (bail recommendation for the theft vignette), with N=275 per arm, 80% power, and α=0.05 (two-tailed), the minimum detectable effect is approximately 8.5 percentage points assuming a baseline pretrial detention rate of 50%. For the recidivism belief outcome (0–10 scale), assuming a standard deviation of 2.5, the MDE is approximately 0.6 scale points with the same parameters.