Description
Fit a generative model of decision-making to each participant's choices or eye movements to estimate personal model parameters (a computational phenotype) and compare rival explanations of behaviour.
The participant is assumed to act on a subjective generative model of the task. The scientist builds an objective model whose likelihood is the probability of the participant's observed actions when that subjective model, set with candidate parameters, is solved on the same stimuli, with a softmax temperature for unexplained randomness. Inverting it gives posterior estimates of each person's priors and preferences (their computational phenotype) and the evidence for alternative models. Individual estimates go to a second level with parametric empirical Bayes and Bayesian model reduction, or to association tests such as canonical variates analysis against questionnaires. Face validity is checked by simulating fictive data and identifiability by parameter recovery. Any normative model, such as reinforcement learning, can supply the likelihood. Asking whether behaviour is optimal is meaningless, since some prior always makes it so; the model is a hypothesis to compare.
Attributes
- Discipline
- Task Analysis & Cognitive Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Evaluation
- Data
- Quantitative
- Effort
- Months
Steps
- Collect trial-level behaviour (choices, saccades, response times) together with the exact stimuli available at each choice, plus between-person variables.
- Formulate a task model (for example a POMDP or reinforcement-learning model) as a function from parameters to predicted action probabilities.
- Specify the likelihood by solving the model on the real stimuli and computing the probability of the observed actions with a softmax temperature, and set priors over plausible parameter ranges.
- Before real data, simulate fictive participants to check face validity and run parameter recovery to confirm parameters are identifiable.
- Invert the model per participant (for example variational Laplace) to obtain posterior parameters and model evidence for rival models.
- Analyse at group level with parametric empirical Bayes and Bayesian model reduction, or relate parameters to questionnaires and outcomes with association tests.
Sources
- Active Inference (Parr et al., 2022)
Also known as
- Model-based data analysis and computational phenotyping by meta-Bayesian model inversion
- Computational phenotyping
- Meta-Bayesian inference
- Observing the observer
- Model-based behavioural analysis
Categories
- User Research
- Task Analysis & Cognitive Modelling