Dmytro Perepolkin: Encoding and updating uncertain judgments using hybrid elicitation and quantile-parametrized distributions

Bayesian parametric inference is about updating the prior beliefs on parameters in light of the new data. Elicitation of parameters, requires that the expert understands the model and the role a particular parameter plays in it. An alternative is to elicit the information about the observable quantities (Kadane & Wolfson 1988), commonly associated with elicitation of predictions about the "next observation". The difficulty with the predictive elicitation is that it makes no distinction between the randomness explained by the model (variability) and uncertainty about the parameters within the model, which makes it difficult to update the expert predictions with the data coming from the new observations.

We propose a hybrid elicitation approach, where elicitation of the observable quantities is accompanied by the elicitation of the uncertainty around them. Hybrid elicitation, combining features of predictive and parametric elicitation, can be used to define the prior for a model defined by a quantile-parametrized distribution. In this talk we present an example of the expert-elicited foods consumption distribution used for exposure assessment, and update it with the observations of actual consumption obtained from the food consumption database.

References