[Impact report] Describe high-level implementation choices for propensity score

[chiyc]

Background

A recent support ticket asked for details about our propensity score implementation. The existing documentation for Causal Impact doesn't say how it's calculated. There is a linked whitepaper, but it's sparse on details about the propensity score.

Changes

This adds some findings from an investigation of the propensity modeler service code to answer the following questions:

• How are the 10 subclasses constructed in practice?
• Which model/inputs are used to estimate the propensity score (feature set, handling of missing values, regularization, balancing heuristics)?
• How are weights computed across sub-classes (inverse variance, subclass size, or other)?
• Any minimum sample size or diagnostics that would cause Causal Impact to skip or fail a subclass?

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs	Ready	Preview, Comment	Mar 10, 2026 6:21pm