A Comparison of Agent-Based Models and the Parametric G-Formula for Causal Inference

Area of Science:

• Causal inference
• Epidemiology
• Computational modeling

Background:

• Accurate treatment effect estimation is crucial for clinical decision-making.
• Parametric g-formula and agent-based models (ABMs) are methods used in the absence of randomized trials.
• ABMs are often used for multi-population effect estimation, requiring stronger assumptions than the g-formula.

Purpose of the Study:

• To describe potential biases in agent-based models (ABMs) when their underlying assumptions are violated.
• To compare the performance of ABMs and the parametric g-formula under differing population assumptions.

Main Methods:

• Estimated 12-month mortality risk in simulated populations with varying prevalence of unmeasured confounders.
• Compared causal effect estimates from ABMs and the g-formula using data exclusively from the target population versus data from other populations.

Main Results:

• Both ABMs and the g-formula accurately estimated mortality and causal effects when all input data originated from the target population.
• ABMs produced biased mortality and causal effect estimates when input data came from populations with different distributions of unmeasured outcome determinants.
• Bias in ABMs persisted even when the true causal effect was null.

Conclusions:

• When all data inputs are from the target population, both ABMs and the g-formula can yield valid causal inferences, assuming no unmeasured confounding or model misspecification.
• ABMs are susceptible to bias when extrapolated to populations with different distributions of unmeasured factors, even with an identical causal network.
• Careful consideration of data source and population characteristics is essential for reliable causal inference using ABMs.