Phase 2

The Hypothesis

01

Solving the Problem

Remember the two methodological problems? Different stakes and probabilities. Mismatched uncertainty levels.

The Solution:

Same bet. Different source. If you see the exact same probabilities in both games, any difference in behavior must come from who or what is making the decision—not the math.

The Innovation: Inverted Trust Game

Berg's (1995) Trust Game has a problem: you don't know the probabilities. You just know someone could betray you. That's second-order risk—uncertainty about uncertainty.

My modification gives you explicit probabilities. First-order risk. You see exactly what could happen and how likely it is. The only question: is it a human making that choice, or a computer?

Social Condition (Trust)

Invest 5 tokens with Player MarVis

Amount quadruples → 20 tokens

You see: 65% chance they return 10 tokens, 35% chance they return 0

Non-Social Condition (Risk)

Invest 5 tokens in Lottery XQZRTY

Amount quadruples → 20 tokens

You see: 65% chance you get 10 tokens, 35% chance you get 0

The Critical Point

Identical probabilities. Identical stakes. Identical structure. The only difference is the label: a person's name vs. a nonsense string. If Prospect Theory parameters differ, it's because social context changes how you process risk.

How It Works

300 Trials, Two Conditions

Structure

  • 20 mini-blocks (15 trials each)
  • Alternating Social/Non-Social
  • 150 trials per condition

Parameters

  • Investment: 1–10 tokens
  • Probabilities: 5%–95% (5% steps)
  • Outcomes: 0–39 tokens
  • Multiplication factor: 4

What We're Testing: 16 Nested Models

Prospect Theory has four key parameters: α (outcome sensitivity), λ (loss aversion), γ⁺ (gain probability weighting), δ⁻ (loss probability weighting).

The question: Do these parameters differ between Trust and Risk? We test 16 models—every possible combination of "same" vs. "different" for each parameter.

H1

Betrayal Aversion

Does λtrust ≠ λrisk? Losing money to a person might hurt more than losing to chance.

Expected: λtrust > λrisk (social losses sting harder)
H2

Domain-Specific Outcome Processing

Does αtrust ≠ αrisk? Maybe you're more sensitive to differences when money comes from a person.

Expected: αtrust > αrisk (sharper sensitivity in social contexts)
H3

Probability Weighting Differentiation

Do γ⁺trust ≠ γ⁺risk or δ⁻trust ≠ δ⁻risk? Do you distort probabilities differently when trusting people?

Expected: γ⁺trust > γ⁺risk and δ⁻trust < δ⁻risk

We compare all 16 models using likelihood ratio tests and information criteria (AIC, BIC). The data will tell us which parameters actually differ—and which don't.

Next: Building the Machine

This design is theoretically motivated. But 300 trials with hundreds of possible parameter combinations—how do we pick the most informative trials?

Explore Phase 3: Optimization