Metacognitive Computations for Information Search -- Code

Overview

Welcome! This repository contains the code to reproduce the results from the paper "Metacognitive Computations for Information Search: Confidence in Control", authored by Lion Schulz, Steve Fleming and Peter Dayan. Don't hesitate to get in touch if you have any questions or comments (corresponding author: Lion Schulz).

Code organisation

The code in this repository is organized as follows:

Directly in the directory: Rmarkdowns reproducing the plots in the computations and results section of the paper, and the appendix.
- MetaSearch_01_Introduction.Rmd produces the figures from the introduction
- MetaSearch_02_Results.Rmd produces most theoretic results and the accompanying figures
- MetaSearch_03_Results_Data.Rmd produces additional theoretic results and analyzes the data
- MetaSearch_04_Appendix.Rmd produces the appendix results
In backend/:
- Several global_[...].R files which load the necessary libraries, custom functions and settings for plotting
  - global_all.R calls all of these and just serves as a shortcut for the Rmarkdowns
  - global_functions.R sources or specifies custom functions
  - global_libraries.R loads all the necessary R libraries for the project
  - global_setttings.R defines settings for the plotting in the main Rmarkdowns. This includes colour schemes, line sizes, etc. It also swithces on anti-aliasing in Windows.
- backend/data_analysis: Functions for the data analysis.
- backend/models/: Model code, divided by model class. Functions used by all models are in the joint folder.
- backend/plotting/: Several .R files that specify code for specific figures in the Rmarkdowns.
- backend/hmetad contains the code for fitting meta-d'. This is an adapted version from https://github.com/metacoglab/HMeta-d
output/ is the folder into which data and plots are saved. This contains high resolution model runs.
data/ contains data from Schulz et al. (2020, PNAS) both in unprocessed form as well as with analyses applied. It also contains simulations.

The repository also includes files necessary for the renv environment management package (see more below). These include the renv/ folder, and the .Rprofile, renv.lock files

Models

The code for the two main models is in backend/models/. This code is written to work for the results in the paper. As such, it returns the relevant summary statistic given the relevant parameters. Thus, we have two main files for the two classes of models:

model_Pd.R for the postdecisional model
model_2o.R for the vanilla second-order model
- model_2o_subjective.R adapts the second-order model to allow for over- and underconfidence

Both functions rely on model_functions_joint.R. The second-order model also uses model_2o_functions.R which contains a ported version of the original second-order confidence.

Reproducing the figures

(1) Download Github directory

First, download the entire Github repository and R project and not just a single file. There are dependencies between folders and the environment within the project which is itself in a managed environment.

(2) Open the R project

This project uses the renv library to control the environment and to make the code more reproducible. To use this, you first need to launch the R project MetaSearch_paper.Rproj using RStudio. Then, renv should automatically bootstrap itself, thereby downloading and installing the appropriate version of renv into the project library.

After this has completed, you then use renv::restore() to restore the libraries used in the project locally on your machine. You can find more information on renv here and on the collaboration functionality used here. Note that renv takes care of the libraries (and their versions), but not of the specific version of R used. You will get a pointer upon calling renv::restore() on the R version used here (4.1.0.). For the best possible compatibility, we advise using this specific version of R, although others should also work.

(3) Running the Rmarkdowns

The main code for the figures in the paper is organized in RMarkdowns, where R code blocks ("chunks") are embedded within a file written in markdown. Output produced by these chunks is produced inline in the markdown.

lionschulz/SchulzFlemingDayan