My favorites | Sign in
Project Home
READ-ONLY: This project has been archived. For more information see this post.
Search
for
Pessiglione2006  
reproducing the analysis with the toolbox
Updated Apr 22, 2013 by jean.dau...@gmail.com

Paper

Dopamine-dependent prediction errors underpin reward-seeking behaviour in humans.
Pessiglione M, Seymour B, Flandin G, Dolan RJ, Frith CD.
Nature (2006), 442(7106):1042-5.
PDF supplementary material

Short description of the study

I here focus on the computational modelling of behavioral data, leaving aside model-free and imaging analysis of the study

In a instrumental learning task, subjects are repeatedly asked to choose among two items (binary decision) of one of 3 fixed pairs of alternatives.
The 3 different pairs correspond to 3 different conditions GAIN/LOSS/NEUTRAL characterized by the outcome generation of the alternatives.

  • In the GAIN condition, one item leads to a GAIN with 0.8 and to NOTHING with probability 0.2 while the other has reversed probabilities
  • In the LOSS condition, one item leads to a LOSS with 0.8 and to NOTHING with probability 0.2 while the other has reversed probabilities
  • In the NEUTRAL condition, both items lead to NOTHING

The 3 pairs are equally pseudo-randomly presented to the subject.

Some subjects performed the task under pharmacological conditions. Therefore three groups of subjects can be separeted

  • control (no treatment)
  • L-dopa
  • Haloperidol

Inferences from behavioral data

Here, we won't redo the analysis from the article, but present the kinds of questions that can be adressed:

  1. show that behavioral performance in an intrumental learning task can be accounted for by a reinforcement learning model. (Does a RL model explains data better than a random model?)
  2. show that difference in behavioral performance for normal vs pharmacological conditions can be explained by a parametric modulation in the model. (Does a model with a set of parameter for each group explains the data better than a model that uses same parameters for all groups)
  3. show that treatment of Gains and Losses are different (Does a model with different parameters for gains and losses explains data better than a model that uses same parameters for both conditions

I propose an answer to these questions here.

Models

Given the inferences we want to make, we need to build the following models

  • Model 1 : Single learning rate, single decision noise for the 3 groups
  • Model 2 : One learning rate and one decision noise for each group
  • Model 3 : One learning rate and one decision noise for each of the two conditions (no distinction between groups)

Remarks

In the model we propose, decisions for the different pairs are independent. Each pair can therefore be treated separately as different sessions of the same task.

This is what we will do here, using the Extension to multiple session of the toolbox. Therefore the models we will define only deal with a single pair, which makes things easier.

Note that considering multiple session doesn't necessarily mean that we consider different parameters of the model for the different sessions (this is indeed not what we want here)

Code and results

Code for the following analysis can be in the Example folder of the toolbox.

  • Synthetic data : Simulate_data_Pessiglione2006.m

Here we will use synthetic data. We choose the parameters of our models as those reported as best fits in the article. Those fits are not reported for individual subjects. Only the mean and the 95% confidence interval are reported. We will sample parameters assuming that the distribution is gaussian and matches those.

  • Model defintion : invert_data_Pessiglione2006_MX.m (X=1,2,3)

  • Inversion scripts : g_softmax_Pessiglione2006_2Q_MX.m and f_Qlearn_Pessiglione2006_2Q_MX.m (X=1,2,3)

Powered by Google Project Hosting