When A/B testing and similar methods are difficult to implement, deriving true effects from "observational data" that includes biases due to background factors (confoundings) requires both mathematical understanding and intuitive verification. This presentation introduces an approach that highly addresses the challenges of confounding by utilizing the "Causal Treatment Effect" platform, which is standard in JMP 19.
JMP 19 represents a significant evolution from the previous "SAS Japan Propensity Score Matching Add-in," implementing "IPW," which utilizes all data, and the cutting-edge "AIPW (Dual Robust Method)," which can derive correct causal effects even if one of the two models is incorrect.
This session will unravel the mathematical mechanisms of these methods, explain practical selection criteria, and provide a simple way to switch between effects on the entire population (ATE) and effects on the treatment group (ATT). Furthermore, we will demonstrate the usefulness of visual graph diagnostic tools such as covariate balance plots, which allow for at-a-glance verification of the degree of bias reduction. Based on mathematical theory, we will propose a workflow that enables anyone to intuitively identify "true causal effects" and make sound decisions.
