Name: Expanding the use of Bayesian Models – Rethinking+Stan
Start: 2025-05-16T04:00:01-0500
End: 2025-05-16T04:00:01-0500

Friday May 16, 2025 TBA

One of the great things about the Bayesian framework is that it allows you to more easily create and think about models that are structured as “how you think the world works” as opposed to some structure imposed by software. But, what if the model you want to estimate doesn’t fit in the model framework that commercial software imposes? In this case you are frequently limited to various general purpose MCMC software packages (e.g. WinBugs, and more recently, Stan). These programs are flexible but tend to be fragile. By fragile I mean that they tend to throw error messages that stop processing and provide little insight into what is wrong with the model.

Enter “Rethinking Statistics” by Richard McElreath. Dr. McElreath has a wonderful online set of lectures that go along with his book “Rethinking Statistics” that provides an easy to understand step by step introduction to modeling in the Bayesian framework. In addition, his “rethinking” package in R allows practitioners to build models that are then translated into Stan programs for fitting and analysis. The “rethinking” package also provides a number of useful tools for extracting posterior distributions from the fitted model for analysis. Overall, the workflow has many fewer roadblocks for the practitioner with limited experience.

In this talk, I walk attendees through some of the basic steps needed to use the “rethinking” package and Stan. Utilizing an example of a standard taste test study where consumers taste and evaluate 4 of 9 products I will show how to construct and fit a model that evaluates and adjusts for order bias and scale usage bias. It is well known in taste tests (and Olympic judging) that there is bias in the evaluation that is due to the order seen. It is also well known in Market Research that there are “high raters” and “low raters” in the population of consumers. In a taste test, where sample size is limited, these biases can cause misleading results. By using Bayesian methods we can not only adjust for these biases but we can easily understand the magnitude of their effects without being limited to the small set of consumers who saw product A in the first position.

This talk will include detailed code and live demonstration of the approach.

Speakers

Michael Conklin

Principal Consultant, 56Stats LLC

For the past 30+ years, Michael has led advanced analytics teams at GfK (Now NIQ) and MarketTools (now MetrixLab). He has constantly slaked his thirst for knowledge by expanding the repertoire of models used to solve marketing and research problems.

Friday May 16, 2025 TBA

4 - Mostly Technical

Data Tech 2025

Michael Conklin

Log in to save this to your schedule, view media, leave feedback and see who's attending!