Bayesian implementation of Targeted Maximum Likelihood Estimation for uncertainty quantification in causal effect estimation

TL;DR

This paper introduces three Bayesian variants of the Targeted Maximum Likelihood Estimation (TMLE) method to quantify uncertainty in causal effect estimates, demonstrating improved performance especially with small datasets.

Contribution

The paper develops and compares three Bayesian TMLE approaches, including a novel Bayesian network implementation, for probabilistic causal inference.

Findings

BN-TMLE outperforms classical methods in small data scenarios.

Bayesian approaches provide probabilistic uncertainty quantification.

Performance is comparable to classical methods with large datasets.

Abstract

Robust decision making involves making decisions in the presence of uncertainty and is often used in critical domains such as healthcare, supply chains, and finance. Causality plays a crucial role in decision-making as it predicts the change in an outcome (usually a key performance indicator) due to a treatment (also called an intervention). To facilitate robust decision making using causality, this paper proposes three Bayesian approaches of the popular Targeted Maximum Likelihood Estimation (TMLE) algorithm, a flexible semi-parametric double robust estimator, for a probabilistic quantification of uncertainty in causal effects with binary treatment, and binary and continuous outcomes. In the first two approaches, the three TMLE models (outcome, treatment, and fluctuation) are trained sequentially. Since Bayesian implementation of treatment and outcome yields probabilistic predictions, the first approach uses mean predictions, while the second approach uses both the mean and standard deviation of predictions for training the fluctuation model (targeting step). The third approach trains all three models simultaneously through a Bayesian network (called BN-TMLE in this paper). The proposed approaches were demonstrated for two examples with binary and continuous outcomes and validated against classical implementations. This paper also investigated the effect of data sizes and model misspecifications on causal effect estimation using the BN-TMLE approach. Results showed that the proposed BN-TMLE outperformed classical implementations in small data regimes and performed similarly in large data regimes.