Bayesian Causal Uplift

Bayesian
DAGs
Causal Inference
PyMC
Estimating heterogeneous treatment effects (CATE) from observational engagement logs using DAGs and PyMC.
Published

September 7, 2026

Code & Infrastructure: Full implementation, Docker environment, and tests are available on GitHub: github.com/krystiankorzec/bayesian-causal-uplift

In high-volume digital products and lifecycle marketing, randomized experiments (RCTs) are not always available or feasible. Whether evaluating the impact of past promotional campaigns, analyzing user-selected communication frequency, or working with observational feature logs, naive regression models frequently mislead decision-making.

When highly active users naturally self-select into higher message exposures or feature usage, standard ML models learn correlations that severely overestimate the true incremental uplift.

This post demonstrates how combining causal Directed Acyclic Graphs (DAGs) with Bayesian G-computation (standardization) isolates true Conditional Average Treatment Effects (CATE) from observational telemetry while providing full posterior uncertainty for policy decisioning.

The Confounding Trap in Engagement Logs

Consider a common scenario: evaluating whether sending an additional targeted communication (\(A \in \{0, 1\}\)) increases user conversion (\(Y\)).

In historical logs, highly engaged users (\(X\)) are significantly more likely to receive or open communications (\(A=1\)) and are also naturally more likely to convert (\(Y=1\)) regardless of the message.

\[\begin{aligned} X &\longrightarrow A \\ X &\longrightarrow Y \\ A &\longrightarrow Y \end{aligned}\]

If we fit a standard supervised model to predict \(Y\) given \(A\) and \(X\), the model estimates the conditional expectation \(E[Y \mid A=1, X]\). However, for decision-making and policy optimization, we require the interventional expectation \(E[Y(A=1) \mid X]\)—what would happen if we forced the treatment across the population.

The Solution: Bayesian G-Computation

G-computation (G-formula standardization) bridges observational telemetry and counterfactual inference in three steps:

  1. Parametric Outcome Modeling: Fit a statistical model \(P(Y \mid A, X, \theta)\) over the observed data.
  2. Counterfactual Simulation: For every user in the target dataset, clone the covariate matrix \(X\) twice:
    • Force \(A = 1\) to simulate the treated outcome \(Y(1)\)
    • Force \(A = 0\) to simulate the untreated outcome \(Y(0)\)
  3. Marginalization & CATE Estimation: Compute the posterior distribution of incremental uplift:

\[\text{CATE}(x) = E[Y \mid A=1, X=x, \theta] - E[Y \mid A=0, X=x, \theta]\]

By integrating Bayesian priors into the outcome model (e.g., using PyMC), we don’t just get a point-estimate of CATE—we obtain a full posterior distribution over the incremental uplift, giving us built-in safety guardrails against negative treatment effects.