This event has already taken place. Slides for the different talks are available below.
Douwe van der Schaaf: Pointlogic
Quentin F. Gronau: Generic Marginal Likelihood Estimation for Stan Models
Alina Ferecatu: Hierarchical Bayesian Analysis – From a binary logit to advanced models of bounded rationality
Tuesday, 27 November 2018.
Fascinatio Boulevard 290
3065 WB Rotterdam
The building is called “De Mark” and the meetup will take place on the 11th floor. A route description can be found here.
18:00 – 18:30. Pizzas (vegetarian and non-vegetarian) made available by Pointlogic
18:30 – 18:50. Welcome & presentation Pointlogic
18:50 – 19:35. First speaker: Quentin Gronau
19:35 – 19:50. Break
19:50 – 20:35. Second Speaker: Alina Ferecatu
20:35 – 21:30. Drinks
You can register here.
Pointlogic, 2 years ago acquired by multinational Nielsen, is a media research company. A part of Pointlogic’s business consists of analysing the effect of advertising on brand KPIs or TV program tune-in. For this purpose, Pointlogic uses (hierarchical) Bayesian regression models applied to respondent-level data. Stan is used to estimate the parameters of the Bayesian models. Currently, efforts are being made to run Stan in parallel in the cloud.
Generic Marginal Likelihood Estimation for Stan Models
Quentin F. Gronau, Henrik Singmann, Eric-Jan Wagenmakers
In many statistical applications, the question of interest can be formulated as a model comparison between two or more possibly non-nested models. A Bayesian solution is to compute posterior model probabilities and Bayes factors. To obtain these quantities one needs to evaluate a model’s marginal likelihood which is often a high-dimensional integral that cannot be solved analytically. In this talk, we discuss bridge sampling, a sampling-based method for estimating the marginal likelihood, and we demonstrate how the bridgesampling R package can be used to obtain estimates of the marginal likelihood for models fitted in Stan in an automatic fashion.
Hierarchical Bayesian Analysis – From a binary logit to advanced models of bounded rationality
The presentation will walk you through how to iteratively develop and estimate a model using stan. The decision problem I will focus on is a behavioral task in which subjects learn the best outcome in experimental games. To illustrate the task, I will play a game involving a multiarmed bandit problem.
To model players’ behavior, I will start with a binary logit model, build it to a hierarchical model, and then introduce a quantal response specification which allows us to model players’ learning patterns. As a last step, I will build a customized likelihood function to assess decision makers’ learning behavior and other relevant behavioral patterns.
On the technical side, the first part of the presentation will be a step-by-step guide to estimating a simple Bayesian model in Stan. I will then show how to estimate hierarchical models with non-centered reparameterization. I will also briefly discuss model checking and model comparison using the loo package.