TL;DR
This paper introduces CHaRTr, an R package designed to simplify the complex process of modeling decision-making behavior. It provides a unified framework for implementing, fitting, and comparing a wide range of sophisticated decision-making models, making advanced quantitative analysis more accessible to researchers.
What Problem Does This Paper Address?
Decision-making is a fundamental cognitive process, and researchers have developed numerous mathematical models to describe how we make choices and how long it takes us. However, applying these models, especially newer ones that incorporate concepts like "urgency" or "collapsing boundaries," requires significant expertise in statistics, programming, and specialized software. Existing toolboxes often focus on a single model (like the diffusion decision model) and lack the flexibility to easily compare a broad spectrum of competing theories. This gap hinders researchers from rigorously testing and advancing our understanding of decision-making.
Methods
CHaRTr offers a comprehensive R-based framework for decision-making model analysis. Key features include:
- Model Specification: Users can define new models by writing C code, which is then compiled into a shared library that R can load. This allows for flexibility in implementing novel theoretical constructs.
- Parameter Estimation: The package employs differential evolution, a robust optimization algorithm, to estimate model parameters from behavioral data. It also supports simulation-based estimation for models lacking closed-form solutions.
- Model Selection: CHaRTr provides tools to compare models using information criteria like AIC and BIC, and calculates model weights to quantify the relative support for each model.
- Visualization: Quantile probability (QP) plots are used to visualize both observed and predicted choice probabilities and response time distributions, offering a compact way to assess model fit.
Datasets
The paper uses simulated data generated from various decision-making models to demonstrate CHaRTr's capabilities. Additionally, it applies CHaRTr to analyze publicly available behavioral data from two monkeys performing a random-dot motion decision-making task (Roitman and Shadlen, 2002).
Key Insights and Ideas
- Democratizing Advanced Modeling: CHaRTr significantly lowers the barrier to entry for using sophisticated decision-making models, enabling a wider range of researchers to conduct rigorous quantitative analyses.
- Model Comparison is Crucial: The paper emphasizes that model selection is best used to guide future research and identify promising hypotheses rather than definitively declaring a single "best" model.
- Flexibility for Future Models: The design of CHaRTr, particularly its C-based model specification, makes it highly extensible, allowing researchers to easily incorporate new theoretical developments.
- Urgency and Collapsing Boundaries Matter: The case studies highlight that models incorporating urgency signals and collapsing boundaries often provide a better account of behavioral data compared to simpler diffusion models.
Key Results
- CHaRTr successfully recovers simulated data-generating models and their parameters, demonstrating its reliability for model recovery analyses.
- In case studies with hypothetical data, CHaRTr correctly identified the data-generating model (DDMS√St and bUGMSv) as the most plausible among a set of candidates.
- Analysis of monkey data (Roitman and Shadlen, 2002) showed that models incorporating urgency signals and/or collapsing boundaries provided a better fit than simpler diffusion models. Variants of the Urgency Gating Model (UGM) were particularly effective.
- The package allows for the simulation of data from over 20 different decision-making models, facilitating hypothesis testing and understanding model predictions.
Limitations and Future Work
- Current Limitations: The current implementation of CHaRTr is primarily designed for tasks where sensory evidence is constant over time and varies along a single dimension. It also assumes independent observers.
- Future Directions: The authors propose several extensions:
- Incorporating "race models" where choices are made by multiple evidence accumulators.
- Developing hierarchical Bayesian methods for parameter estimation to provide full posterior distributions and improve statistical power, especially with limited trials.
- Extending the framework to handle time-varying stimuli and multisensory integration.
- Allowing more model parameters to vary with experimental factors beyond just drift rate.
- Exploring approximate Bayesian techniques for parameter estimation.
Summary updated