Aurevia 2×2 Experiment: AI Authorship, Governance Review, and Investor Allocation Dataset
Description
This dataset contains anonymized, analysis-ready data from a 2 × 2 randomized online experiment examining how AI authorship and formal governance review influence investor responses to earnings communication. The experiment was administered through Qualtrics and used a standardized earnings update for Aurevia Technologies, a fictional publicly listed enterprise-software company. Financial information was identical across conditions; only the disclosed authorship cue (CEO-prepared versus AI-generated) and governance-review cue (no formal review disclosed versus formal review disclosed) varied. The final analytic sample comprises 636 investment-experienced adults drawn from 702 completed randomized responses. The sample-flow documentation reports 66 sequential exclusions based on eligibility and data-quality criteria applied before hypothesis testing. The primary behavioral outcome is the hypothetical amount allocated to Aurevia common stock from a constant-sum $10,000 portfolio. The dataset also includes NonEquityAllocation_USD, defined as bond allocation plus risk-free allocation, equivalently $10,000 minus common-stock allocation. Bond and risk-free allocations are retained as descriptive components of the original portfolio task and should not be interpreted as independent primary outcomes or as direct measures of overall willingness to invest in the company. Additional variables include perceived accountability, perceived governance oversight, trust and credibility, expected stock return, perceived stock risk, investment attractiveness, decision confidence, investment experience, financial knowledge, risk tolerance, AI familiarity, and general trust in AI. The workbook contains four sheets: README, Clean_Data, Codebook, and Sample_Flow. The main analyses follow an intent-to-treat approach and retain all 636 participants satisfying the pre-outcome eligibility and quality criteria. Manipulation checks were administered after the primary allocation decision and were not used to determine inclusion. Aggregate manipulation-check pass rates are reported by experimental condition, but genuine respondent-level check indicators are unavailable and have not been reconstructed. Participant-level age, direct identifiers, recruitment-channel indicators, and underlying scale-item responses are not included in the preserved analytic dataset. Exact numeric cutoffs for certain timing-based exclusions were also not preserved and are identified transparently in the documentation. The data support reproducibility of the reported analyses but should not be interpreted as evidence of actual investment or trading behavior.
Files
Steps to reproduce
1. Open `Aurevia_2x2_Depo_Clean.xlsx` and import the `Clean_Data` sheet into statistical software. The sample should contain 636 observations and 22 variables. 2. Verify the experimental coding: * `Authorship_AI`: 0 = CEO-prepared; 1 = AI-generated. * `GovernanceReview`: 0 = no review; 1 = formal review. * `AIxReview`: `Authorship_AI × GovernanceReview`. Condition sizes should be 158 for `CEO_NoReview`, 163 for `CEO_Review`, 154 for `AI_NoReview`, and 161 for `AI_Review`. 3. Verify that, for every participant: * `StockAllocation_USD + BondAllocation_USD + RiskFreeAllocation_USD = 10,000`. * `NonEquityAllocation_USD = BondAllocation_USD + RiskFreeAllocation_USD`. * `TotalAureviaAllocation_USD = StockAllocation_USD + BondAllocation_USD`. Common-stock allocation is the primary behavioral outcome. Bond and risk-free allocations are descriptive portfolio components. 4. Calculate condition means for common-stock, bond, and risk-free allocations, trust and credibility, accountability, and governance oversight. 5. Estimate the baseline OLS model with HC3 robust standard errors: `StockAllocation_USD = Authorship_AI + GovernanceReview + AIxReview`. 6. Estimate the adjusted model by adding: * `InvestmentExperience_1to5`; * `FinKnowledgeScore_0to3`; * `RiskTolerance_1to7`; * `AIFamiliarity_1to7`; and * `GeneralTrustAI_1to7`. 7. Calculate these planned contrasts: * AI/no review minus CEO/no review; * CEO/review minus CEO/no review; * AI/review minus AI/no review; * AI/review minus CEO/review; and * `(AI_Review − AI_NoReview) − (CEO_Review − CEO_NoReview)`. 8. Calculate correlations between accountability and trust, governance oversight and trust, and trust and common-stock allocation. Estimate a mechanism-adjusted model including the experimental factors, accountability, oversight, trust, and the five controls. These analyses are associational because perceptions were measured after allocation. 9. Estimate separate factorial models for `ExpectedReturnPct`, `PerceivedStockRisk_1to7`, and `InvestmentAttractiveness_1to7`. Do not treat total Aurevia, bond, or risk-free allocation as independent primary outcomes. 10. Use `Sample_Flow` to reproduce the sample construction: 702 completed responses, 66 sequential exclusions, and final N = 636. 11. Report manipulation checks descriptively from the aggregate `Sample_Flow` counts. Do not create respondent-level indicators or conduct a 539-participant analysis because genuine participant-level check variables are unavailable. The intent-to-treat analysis retains all 636 eligible observations.
Institutions
- Ca' Foscari University of VeniceVeneto, Venice