/COMPASS

Repository for the COMPASS toolbox, created by Maud Beeckmans, Pieter Huycke, Tom Verguts and Pieter Verbeke

Primary LanguagePython

Uni Date

COmputational Power Analysis using Simulations "COMPASS" toolbox

This toolbox has been developed to evaluate statistical power when using parameter estimates from computational models.

In the current version, use is limited to the Rescorla-Wagner (RW) model in two-armed bandit tasks.

More details can be found in the manuscript: https://10.31234/osf.io/dexyk

We now also provide a script to easily fit the RW model on your own empirical data!!!

More details are provided at the end of this page.

Installation guide

Step 1: Downloading all code and storing them locally on your own pc.

Step 2: Creating the PyPower environment

Normally, power analyses with COMPASS should be possible in a basic Python environment.
Nevertheless, to control for version issues, we provide environment files for windows and mac users.

  • Install Anaconda 3 by following their installation guide
  • When the installation is complete, open an Anaconda prompt
  • Go to the directory where the COMPASS files are stored using cd
  • For windows, run: conda env create --file environment_windows.yml and for mac, run conda env create --file environment_mac.yml
  • Allow the installation of all required packages

The model and task currently implemented in COMPASS

The RW model

To specify by the user: meanLR, sdLR, meanInverseTemperature and sdInverseTemperature

The RW model is used to fit participants’ behaviour in this task.
The core of the model is formed by the delta-learning rule and the softmax choice rule.
The model has two free parameters: the learning rate and the inverse temperature (see manuscript for details).
The population distribution of these parameters can be specified by the user in the csv files by completing meanLR, sdLR, meanInverseTemperature and sdInverseTemperature.

Two-armed bandit task

To specify by the user: ntrials, nreversals, reward_probability

Based on the parameters that are implemented in the csv files, a task design is created.
In this task design, there are two stimuli/bandits and two possible actions (selecting bandit 1 or 2).
One bandit is more optimal since it has a higher probability of reward (specified by reward_probability).
For simplicity, we implement that Pr(Reward | optimal_bandit) = 1- Pr(Reward | suboptimal_bandit).
As in classic reversal learning tasks, we provide the option that the identity of the optimal bandit can reverse. Here, one has to specify the frequency of rule reversals (nreversals). If set to zero, there are no reversals.
The design is created for ntrials. As demonstrated in the manuscript, this is a crucial variable for obtaining reliable parameter estimates and high statistical power.

Important note: the required computational time.

To specify by the user: npp, nreps, full_speed

As we perform parameter estimations for nreps Monte Carlo repetitions, computational time can increase exponentially.
The computational time strongly depends on the number of participants (npp) and the number of Monte Carlo repetitions (nreps). We recommend to set nreps to 250. Smaller numbers can be used as well but then power computations will be less precise.
Notably, also increasing the number of trials in the task design (ntrials) can significantly increase the power computation time in COMPASS.

As a partial solution for this computation time, the option is included to run the power analysis on multiple cores. This happens when the user defines the full_speed option as 1; if this option is activated, all minus two cores on the computer will be used for power computations.

When running COMPASS it will asap provide an estimate of how long it will take to calculate the power for each power computation (each row of the csv files specify one power computation).
This estimate is based on the time it takes to execute a single repetition and calculated by multiplying the total number of repetitions by the time required for a single repetition, divided by the number of cores that are used in the power analysis.

If you want to stop the process whilst running, you can use 'ctrl + C' in the anaconda prompt shell. This will stop the execution of the script.

Runnig power computations with COMPASS

As described in the manuscript, three criteria for power computations are specified.
For each criterion (IC, EC or GD), we provide a csv file which holds the power variables that should be specified by the user.
For all criteria, power is specified as

power = Pr(Statistic > cut-off | Hypothesis)

Here, the statistic differs across criteria and the cut-off and hypothesis should be specified by the user.

Power computations consist of the following five steps:

  1. Sample npp participants from the population.
    This sampling process is guided by the hypothesis that is specified by the user in the csv files. (population distribution of parameter values (for IC), true correlation (for EC) or difference between groups (for GD))
  2. Simulate data for each participant.
  3. Estimate the best fitting parameters for each participant given the simulated data.
    These are the β€˜estimated parameters’.
  4. Compute statistics.
    The statistic differs across criteria.
    • internal_correlation: correlation between sampled and estimated parameter values.
    • external_correlation: correlation between estimated parameter values and external measure (e.g., questionnaire score).
    • group_difference: T-statistic of difference in parameter values between two groups.
  5. Evaluate which proportion of statistics reached the cut-off value.

The steps for the user

  1. Make sure that COMPASS is installed correctly (see Installation guide above).

  2. Choose a criterion and specify variables in the corresponding csv file.
    Notice that multiple rows can be specified in the csv files, power computations will be performed for each row that is completed by the user

2a) Internal Correlation (IC): Correlation between sampled and estimated parameter values.

Open the InputFile_IC and specify

  • ntrials: integer πœ– [5, +∞[ number of trials within the experiment (minimal 5)
  • nreversals: integer πœ– [0, ntrials[ number of rule reversals within the experiment
  • npp: integer πœ– [5, +∞[ total number of participants within the experiment (minimal 5)
  • meanLR: float πœ– [0, 1] mean of the assumed population distribution of learning rates
  • sdLR: float πœ– [0, 1] sd of the assumed population distribution of learning rates
  • meanInverseTemperature: float πœ– [0, 1] mean of the assumed population distribution of inverse temperatures
  • sdInverseTemperature: float πœ– [0, 1] sd of the assumed population distribution of inverse temperatures
  • reward_probability: float πœ– [0, 1] The probability of reward for the optimal bandit in the two-arm bandit task.
  • tau: float πœ– [0, 1] the value against which the obtained statistic will be compared to define significance of the repetition.
    • correlation: cut_off = minimally desired correlation - recommended: 0.75
  • full_speed: integer (0 or 1) Define whether you want to do the power analysis at full speed.
    • 0 = only one core will be used (slow)
    • 1 = (all-2) cores will be used (much faster, recommended unless you need your computer for other intensive tasks such as meetings)
  • nreps: integer πœ– [1, +∞[ Number of repetitions that will be conducted to estimate the power
    • Recommended number: 250
  • output_folder: string Path to the folder where the output-figure(s) will be stored
    • e.g. "C:\Users\maudb\Downloads"

2b) External Correlation (EC): Correlation between estimated parameter values and external measure (e.g., Questionnaire scores).

Open the InputFile_EC and specify

  • ntrials: integer πœ– [5, +∞[ number of trials within the experiment (minimal 5)
  • nreversals: integer πœ– [0, ntrials[ number of rule reversals within the experiment
  • npp: integer πœ– [5, +∞[ total number of participants within the experiment (minimal 5)
  • meanLR: float πœ– [0, 1] mean of the assumed population distribution of learning rates
  • sdLR: float πœ– [0, 1] sd of the assumed population distribution of learning rates
  • meanInverseTemperature: float πœ– [0, 1] mean of the assumed population distribution of inverse temperatures
  • sdInverseTemperature: float πœ– [0, 1] sd of the assumed population distribution of inverse temperatures
  • reward_probability: float πœ– [0, 1] The probability of reward for the optimal bandit in the two-arm bandit task. *True_correlation: float πœ– [-1, 1] The hypothesized correlation between the learning rate parameter values and the external measure
  • TypeIerror: float πœ– [0, 1] The allowed probability to make a type I error; the significance level
    • standard (and recommended) value: 0.05
    • correlation: cut_off = minimally desired correlation - recommended: 0.75
  • full_speed: integer (0 or 1) Define whether you want to do the power analysis at full speed.
    • 0 = only one core will be used (slow)
    • 1 = (all-2) cores will be used (much faster, recommended unless you need your computer for other intensive tasks such as meetings)
  • nreps: integer πœ– [1, +∞[ Number of repetitions that will be conducted to estimate the power
    • Recommended number: 250
  • output_folder: string Path to the folder where the output-figure(s) will be stored
    • e.g. "C:\Users\maudb\Downloads"

2c) Group Difference (GD): T-statistic for difference between estimated parameter values of two groups.

Open the InputFile_GD and specify

  • ntrials: integer πœ– [5, +∞[ number of trials within the experiment (minimal 5)
  • nreversals: integer πœ– [0, ntrials[ number of rule reversals within the experiment
  • npp_group: integer πœ– [5, +∞[ number of participants within the experiment (minimal 5)
  • meanLR_g1: float πœ– [0, 1] mean of the assumed population distribution of learning rates for group 1
  • sdLR: float πœ– [0, 1] sd of the assumed population distribution of learning rates for group 1
  • meanLR_g2: float πœ– [0, 1] mean of the assumed population distribution of learning rates for group 2
  • sdLR_g2: float πœ– [0, 1] sd of the assumed population distribution of learning rates for group 2
  • meanInverseTemperature_g1: float πœ– [0, 1] mean of the assumed population distribution of inverse temperatures for group 1
  • sdInverseTemperature_g1: float πœ– [0, 1] sd of the assumed population distribution of inverse temperatures for group 1
  • meanInverseTemperature_g2: float πœ– [0, 1] mean of the assumed population distribution of inverse temperatures for group 2
  • sdInverseTemperature_g2: float πœ– [0, 1] sd of the assumed population distribution of inverse temperatures for group 2
  • reward_probability: float πœ– [0, 1] The probability of reward for the optimal bandit in the two-arm bandit task.
  • TypeIerror: float πœ– [0, 1] The allowed probability to make a type I error; the significance level
    • standard (and recommended) value: 0.05
  • full_speed: integer (0 or 1) Define whether you want to do the power analysis at full speed.
    • 0 = only one core will be used (slow)
    • 1 = (all-2) cores will be used (much faster, recommended unless you need your computer for other intensive tasks such as meetings)
  • nreps: integer πœ– [1, +∞[ Number of repetitions that will be conducted to estimate the power
    • Recommended number: 250
  • output_folder: string Path to the folder where the output-figure(s) will be stored
    • e.g. "C:\Users\maudb\Downloads"

  1. Run the PowerAnalysis.py script using the correct Anaconda 3 environment.

    If one followed the Installation guide above, a PyPower environment has been created.

    To use this environment:

    • Open Anaconda prompt
    • Now, run: conda activate pyPower

    To run COMPASS:

    • Go to the directory where the COMPASS files are stored using cd
    • Now, run: python PowerAnalysis.py IC, python PowerAnalysis.py EC or python PowerAnalysis.py GD depending on the criterion that you want to use.

  2. Check the output in the shell & the stored figure(s) in the output_folder

    • power estimate: the probability to obtain adequate parameter estimates.
    • probability density plot of the Statistic of interest: a plot visualising the obtained values for the Statistic of interest in all power recovery analyses
      • x-axis: values for the statistic of interest (correlation or T-Statistic)
      • y-axis: probability density for each value

    Example output (EC criterion):

    image image

Fitting the RW model on your data

Requires a specifically structured folder with behavioural data files:

  • Relevant files should follow this format: Data_Subject_{SubjectID}.csv Here, {SubjectID} can be a number or string identifying that specific subject.
  • Files should contain four columns: Trial, Stimulus, Response and Reward
  • Stimulus and Response should be coded as integer going from 0 to the number of stimuli or responses (e.g., 0 and 1 if there are two response options)
  • No other csv files should be in your folder!

Returns:

  • Fitting_results.csv file, containing a row for each subject and four columns: SubjectID, Estimated_LR, Estimated_InvTemp and Negative_LogL
  • Simulated files for each subject. A copy of each individual file is saved, containing two additional rows: Response_likelihood and PE_estimate

Running the fitting procedure:

To use our environment:

  • Open Anaconda prompt
  • Now, run: conda activate pyPower

To run the fitting procedure:

  • Go to the directory where the COMPASS files are stored using cd
  • Now, run: python Fit_data.py {"datafolder"} where {"datafolder"} represents a string indicating the path to the folder where your data is located

Contact

Last edit: June 19th 2023