Seeing the Cliff Before You Fall: The Cognitive Cost of Misperception

Participants are randomly assigned to one of four arms in a 2×2 design. The first dimension varies the format in which benefit information is presented: (1) an interactive dashboard that visualizes how total household resources change at different income levels, or (2) an AI assistant that delivers the same underlying information conversationally and maps program rules to the participant's specific household. The second dimension independently varies whether participants receive recovery information — a brief statement indicating that workers who experience a benefit cliff typically return to their prior resource level within 1–3 years as wage growth offsets benefit losses. All participants first complete a baseline stage with no additional information beyond a standard text description of the scenario, replicating the information environment workers encounter in real life. In the second stage, participants receive their assigned treatment and answer the same questions again. This design allows us to separately identify the effect of information format (Frictions 1 and 2: computational complexity and situational mapping) from the effect of reducing perceived permanence of losses (Friction 3: intertemporal uncertainty).

Participants are recruited via Prolific and must meet two eligibility criteria: household income between 100–250% of the Federal Poverty Level and current or past receipt of at least one safety-net program (SNAP, Medicaid, TANF, CCDF, ACA subsidies, or EITC). After completing an intake survey capturing household composition, employment, and benefit receipt, participants are matched to one of three vignette templates (single adult no children, single parent with children, two-adult household with children) based on their reported household structure.
In Stage 1 (baseline), all participants read a text description of a hypothetical worker facing a job offer that increases earnings near a benefit cliff threshold. The description replicates the kind of fragmented, agency-style program information workers encounter in real life. Participants report (1) their predicted change in total monthly household resources if the worker accepts the offer and (2) whether they would recommend accepting the offer.
In Stage 2, participants receive their randomly assigned treatment. Arms A and B receive access to an interactive Policy Rules Database (PRD) dashboard that visualizes net household resources across income levels. Arms C and D interact with an AI assistant (Claude, Anthropic) that provides the same underlying PRD data conversationally, personalized to the vignette household. Arms B and D additionally receive a recovery information box stating that workers in similar occupations typically return to their prior resource level within a specified number of years, calculated from sector-specific wage growth rates. Arms A and C receive no recovery information. Participants repeat the same two questions from Stage 1.
In Stage 3, all participants see the actual resource change computed from PRD rules alongside their Stage 1 and Stage 2 estimates. They then answer questions on reconsideration of the job decision, tolerance for short-term loss, and willingness to pay for benefit information. Accuracy bonuses of $1.50 per stage are awarded when the participant's predicted monthly resource change falls within $300 of the true value.

2026-04-22

2026-04-29

The primary outcome is belief improvement — the reduction in absolute prediction error between Stage 1 and Stage 2, measured in dollars per month:
BI = |ε₁| − |ε₂| = |Δ̂₁ − Δ*| − |Δ̂₂ − Δ*|
where Δ̂ₛ is the participant's stated prediction of monthly household resource change at stage s and Δ* is the true change computed from Policy Rules Database rules. A positive value indicates the treatment brought beliefs closer to the truth. This outcome captures the extent to which each information format reduces misperception of benefit cliff magnitudes.

Belief improvement is constructed as follows. At Stage 1, the participant reports a predicted monthly resource change Δ̂₁ (a signed dollar amount). At Stage 2, after receiving the assigned treatment, the participant reports a revised prediction Δ̂₂. The true monthly resource change Δ* is computed programmatically from Policy Rules Database (PRD) rules for the vignette household, taking into account wages, taxes, and all relevant benefit programs (SNAP, Medicaid, CCDF, TANF, EITC, Section 8) at both the pre- and post-offer income levels. The absolute prediction error at each stage is |Δ̂ₛ − Δ*|. Belief improvement is the difference |ε₁| − |ε₂| = |Δ̂₁ − Δ*| − |Δ̂₂ − Δ*|, measured in dollars per month. Positive values indicate improvement; negative values indicate that the treatment worsened accuracy. To limit the influence of extreme responses, predicted changes will be winsorized at the 1st and 99th percentiles of the Stage 1 distribution prior to computing errors. The winsorization thresholds will be determined from the data and applied symmetrically to both stages.

Decision quality — an indicator equal to 1 if the participant's job acceptance recommendation aligns with the financially optimal decision (accept if Δ* > 0, reject if Δ* < 0), measured at Stage 1 and Stage 2 separately. The treatment effect is the change in decision quality between stages: Optimal₂ − Optimal₁.
Money-metric regret — the dollar value of welfare loss from a suboptimal decision, defined as Regret = R_best − R_chosen ≥ 0, where R_j is total monthly household resources under option j. Regret is positive only when the participant recommends the dominated option.
Willingness to pay for benefit information — the percentage of the monthly wage increase the participant would pay for a service that reveals exact benefit changes before accepting a job offer (Stage 3), converted to dollars using the vignette wage increase.
Reconsideration — an indicator for whether the participant changes their job acceptance recommendation after seeing the true resource change in Stage 3.
Tolerance for short-term loss — the maximum monthly resource loss the participant would accept for a job with significantly better long-term career prospects (Stage 3), measured in dollars per month.
Stated confidence — self-reported confidence in the prediction at Stage 1 and Stage 2, used to test whether treatments raise confidence without improving accuracy (over-confidence check, particularly relevant for the AI assistant arm).

Decision quality is constructed as a binary indicator from the participant's yes/no job acceptance recommendation at each stage. The optimal decision is defined as accepting when Δ* > 0 and rejecting when Δ* < 0. For vignettes where Δ* = 0, decision quality is coded as 1 regardless of the participant's choice. The treatment effect is the within-person change Optimal₂ − Optimal₁, which takes values in {−1, 0, 1}.
Money-metric regret is computed as the absolute difference between total monthly household resources under the optimal option and the chosen option: Regret = |R_best − R_chosen|. For participants who recommend the optimal option, Regret = 0. For those who recommend the dominated option, Regret = |Δ*|. The outcome is bounded below by zero and measured in dollars per month.
Willingness to pay is elicited as a percentage (0–100) in Stage 3 and converted to dollars by multiplying by the vignette monthly wage increase divided by 100. Responses above 100 or below 0 will be recoded as missing.
Reconsideration is a binary indicator equal to 1 if the participant's Stage 3 job acceptance recommendation differs from their Stage 2 recommendation, after having seen the true resource change. This captures preference-based updating distinct from belief-based updating.
Tolerance for short-term loss is directly elicited in Stage 3 as a dollar amount per month. Responses will be winsorized at the 99th percentile. This outcome is used to assess intertemporal preferences independently of misperception.
Stated confidence is elicited on a scale at Stage 1 and Stage 2. An over-confidence indicator is constructed as high stated confidence combined with large absolute prediction error, defined as confidence above the median and |ε| above the median simultaneously. This is used as a robustness check on the AI assistant arm specifically.

This study uses a 2×2 between-subject experiment. Participants are randomly assigned to one of four arms crossing two dimensions: information format (interactive dashboard vs. AI assistant) and recovery information (absent vs. present). All participants complete a baseline estimation task before receiving their assigned treatment. The target sample is N = 300 U.S. adults recruited via Prolific who currently receive or have recently received at least one government assistance program.

The experiment is implemented in oTree and deployed via Heroku. Participants are recruited through Prolific using prescreening filters for U.S. residency and current or past receipt of at least one safety-net program.
After completing an intake survey, participants are matched to a vignette from a pre-constructed bank of 34 scenarios varying by household type (single adult, single parent, two-adult household), state (Pennsylvania, Georgia, Texas), income level, job sector, and benefit bundle (SNAP, Medicaid, CCDF, TANF, Section 8). Vignettes are matched to participants based on household composition and income range reported in the intake survey. True resource changes (Δ*) are pre-computed using the Atlanta Fed Policy Rules Database (PRD) and stored in the vignette bank prior to data collection.
Randomization to the four arms (A: Dashboard / No recovery, B: Dashboard / Recovery, C: AI / No recovery, D: AI / Recovery) is implemented at the individual level within oTree at the start of Stage 2, stratified by vignette type. The AI assistant is powered by the Anthropic Claude API, called from the oTree backend to avoid CORS issues. The dashboard is an embedded iframe of the PRD interactive tool.
The econometric specification is a between-subject OLS regression of belief improvement on treatment indicators, with vignette fixed effects and controls for intake characteristics (household size, state, benefit receipt, financial stress). Heterogeneity analysis will examine whether treatment effects differ by prior experience with benefit cliffs, financial stress, and whether the vignette involves a cliff (Δ* < 0) vs. a non-cliff (Δ* > 0).
Pre-specified robustness checks include: (1) winsorization of predicted changes at the 1st/99th percentiles, (2) ITT analysis using time-on-page logs to identify non-compliers, (3) bounds analysis for differential attrition by arm, and (4) separate analysis for cliff vs. non-cliff vignettes.

Individual-level randomization implemented automatically by the oTree platform at the start of Stage 2. Participants are assigned to one of four arms using a pseudo-random number generator, stratified by vignette type to ensure balance across treatment cells.

Individual participant.

No

300 participants

300 participants

75 participants — Arm A (Dashboard, no recovery information)
75 participants — Arm B (Dashboard, recovery information)
75 participants — Arm C (AI assistant, no recovery information)
75 participants — Arm D (AI assistant, recovery information)

Primary outcome: belief improvement (reduction in absolute prediction error, measured in dollars per month). At N = 300 (150 participants per main-effect comparison group), α = 0.05, power = 0.80, and assuming a between-subject standard deviation of σ = $200/month (Taubinsky & Rees-Jones 2018, upper bound), the minimum detectable effect is $68/month, representing approximately 15–20% of a typical cliff magnitude (Δ* ≈ −$200 to −$600/month). Under the more conservative Abeler & Jäger (2015) anchor (σ = $250/month), the MDE is $85/month. Both are below the pre-specified focal threshold of $100/month. The sample size formula used is N = 2 × ⌈(z₁₋α/₂ + z₁₋β)² × 2σ² / d²⌉ for a two-sided two-sample t-test. Standard errors will be clustered at the individual level.