Loss Aversion in Teen Decision Models

ISEF Category: Behavioral and Social Sciences

Subcategory: Behavioral Neuroscience  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A teen can look risk-seeking in one task and cautious in another. That makes loss aversion hard to pin down, even when the math seems simple. In this project, you can test a published reinforcement-learning model on public choice data and see where it explains behavior well, and where it misses.

What Is It?

Loss aversion means a loss can feel larger than an equal gain. If you would rather avoid losing $10 than win $10, your choices point in that direction. Reinforcement learning is the idea that you update future choices based on feedback, a little like learning which route home usually saves time.

The Iowa Gambling Task (IGT) and Balloon Analogue Risk Task (BART) are common ways to study risky choice. In the IGT, you pick cards with wins and losses. In the BART, you decide how far to push a balloon before cashing out. A hierarchical Bayesian model acts like a group model with built-in individual differences, so you can estimate both the overall pattern and each participant's version of it.

Why This Is a Good Topic

This is a good science fair topic because you can work from public data, clear variables, and a question with real structure. You are not just summarizing a paper, you are testing whether the model still works when you fit it yourself and compare tasks. That gives you practice with cleaning data, fitting Bayesian models, and reading uncertainty the right way.

Research Questions

• How does the estimated loss-aversion parameter change between the IGT and BART datasets?
• What is the effect of adding hierarchical pooling on out-of-sample prediction?
• Does a model with separate gain and loss learning rates fit the data better than a single-rate model?
• To what extent does a stay-switch bias improve prediction of trial-by-trial choices?
• Which model specification best explains the participant-level spread in risk choices?
• How does the posterior for loss aversion change when you compare weak, medium, and strong priors?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Stable internet connection for downloading the public datasets.
• Kaggle account with the IGT and BART CSV files and data dictionary.
• External drive or cloud folder for backups.
• Notebook for planning model choices and recording run notes.
• A quiet workspace for long model runs.

Advanced Materials

• High-memory workstation or campus cluster access.
• External SSD for large posterior draws and backups.
• Second monitor for diagnostics, trace plots, and model comparisons.
• Secure shared storage for de-identified files and versioned outputs.
• Printed or digital data dictionary for the public datasets.
• Headphones or a quiet room for long coding sessions.

Software & Tools

• Python: Runs the data cleaning, model fitting, and plotting code.
• PyMC: Fits hierarchical Bayesian reinforcement-learning models with MCMC.
• ArviZ: Checks convergence, compares models, and plots posterior distributions.
• pandas: Cleans the trial-level CSV files and reshapes participant data.
• JupyterLab: Keeps your code, notes, and figures in one notebook.

Experiment Steps

1. Define the exact model comparison you want to make, then decide whether you are reproducing one task or comparing both tasks.
2. Reshape the raw trial files into a participant-by-trial table with one row per choice.
3. Build a simple baseline model first, then add hierarchical pooling and separate gain-loss terms.
4. Choose your fit checks before you run the model, including posterior predictive checks and held-out prediction.
5. Plan one extension that tests generalization, such as a subgroup split, an alternate prior, or a task comparison.

Common Pitfalls

• Mixing trial order across participants, which destroys the learning signal the model is supposed to learn.
• Treating IGT and BART as the same task, which can make one loss-aversion parameter explain too much.
• Skipping convergence checks, which leaves you with parameter values that look precise but are not stable.
• Comparing only in-sample fit, which rewards flexible models instead of models that predict new choices.
• Ignoring parameter identifiability, which makes gain, loss, and learning-rate terms trade places in the posterior.

What Makes This Competitive

A strong version does more than reproduce the paper. You compare several model forms, test them on held-out choices, and report whether the same loss-aversion pattern appears in both tasks. If you add prior sensitivity checks and clear plots of participant-level differences, your project starts to look like real research, not a class demo.

Project Variations

• Fit the same model to a different public adolescent dataset, then check whether the loss-aversion pattern repeats.
• Compare a simpler single-rate model with a gain-loss split model to see which one predicts choices better.
• Analyze IGT and BART separately, then test whether one shared parameter set can explain both tasks.

Learn More

• PubMed: Search for review articles on adolescent risk taking, loss aversion, and reinforcement learning.
• PubMed Central: Read full-text papers on decision making and behavioral neuroscience.
• NIH RePORTER: Look up funded projects on adolescent behavior and computational psychiatry.
• Kaggle: Find the public IGT and BART datasets, data notes, and starter notebooks.
• PyMC documentation: Learn how to build hierarchical Bayesian models and sample posteriors.
• ArviZ documentation: Check convergence, compare models, and inspect posterior plots.