Abstract:
We are constantly amazed by the new software and analysis techniques in machine learning and data science; specifically their promise to empower decision analysts to integrate data-driven methods with judgment-based analysis. Previously, such attempts were frustrated by the incoherence of conventional statistical methods, but the rise of Bayesian methods in machine learning have created an opening to build powerful new tools, recognizing the common principles both disciplines derive from.
This tutorial will introduce you to several applications in healthcare and tech that bridge the gap between current decision analysis practice and current probabilistic methods in machine learning. Modern tools answer the question of how to quantify decision uncertainties by discovering and modeling relevant data. Our demonstrations include influence diagram & probabilistic graphical models integrated with machine learning tools that produce proper probabilistic results. In this way, the efforts of data science teams can be aligned with the critical decisions a fast-paced decision-making organization faces.
First Talk, Make Better Project Decisions with Data
Deciding the right projects for your Data Science team is crucial for both the growth of your team and to increase the analytical maturity of the teams you support. This talk will go over experience with a few frameworks that lay out how you can better structure the work of your Data Science teams in a way that increases their value within your organization. We discuss the current Data Science alignment challenges that apply to decision making such as arise in high tech product management.
Second Talk, Integrating Decision Analysis and Data Science, The Bayesian Promise
Decision Analysis and Data Science teams often operate at cross-purposes, using distinct frameworks. This talk works through a coding demonstration of how machine learning classification trees can be applied in a principled way to Decision Analysis. Extending a well-known textbook example called "The Used Car Buyer", we go through the steps to build an influence diagram decision model, then learn its probability model from a data set, and finally integrate the two models, including showing how it is possible to tune the predictions to the priors taken from the decision model, and to extend the analysis to include value of information analysis. The result is a novel analytical approach that combines software tools that solve influence diagrams with Bayesian statistical methods. This advances the computational reach of Decision Analysis, aligning it with contemporary evidence-based practices.
Third Talk, Evidence Synthesis for Decision Making in Healthcare
This is an introduction to economic decision modeling, using healthcare examples. We demonstrate how modern statistical tools for Bayesian analysis enable information on the effectiveness of medical interventions to be integrated with economic considerations to provide decision makers with a comprehensive overview of cost effectiveness. Going beyond decision trees based solely on point estimates, we show how both clinical and economic uncertainty can be incorporated in a single computational framework using probability distributions. We introduce some of the modern Markov Chain Monte Carlo (MCMC) tools and show how they foster reproducible research and literate analysis. Attendees will have access to a Quarto notebook enabling them to run all of the R analysis code from the talk.