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

Abstract

Decision-making requires choosing from treatments on the basis of correctly estimated outcome distributions under each treatment. In the absence of randomized trials, 2 possible approaches are the parametric g-formula and agent-based models (ABMs). The g-formula has been used exclusively to estimate effects in the population from which data were collected, whereas ABMs are commonly used to estimate effects in multiple populations, necessitating stronger assumptions. Here, we describe potential biases that arise when ABM assumptions do not hold. To do so, we estimated 12-month mortality risk in simulated populations differing in prevalence of an unknown common cause of mortality and a time-varying confounder. The ABM and g-formula correctly estimated mortality and causal effects when all inputs were from the target population. However, whenever any inputs came from another population, the ABM gave biased estimates of mortality-and often of causal effects even when the true effect was null. In the absence of unmeasured confounding and model misspecification, both methods produce valid causal inferences for a given population when all inputs are from that population. However, ABMs may result in bias when extrapolated to populations that differ on the distribution of unmeasured outcome determinants, even when the causal network linking variables is identical.

Effect heterogeneity and variable selection for standardizing causal effects to a target population

“…If effect homogeneity in distribution holds conditional on V , one can use a third standardization formula (approach 3 in Table 2), based on separately standardizing the conditional risk under treatment and the conditional risk under no treatment from the study population, to the distribution of V in the target population. Murray et al [21] shows how such standardization relates to agent-based models, and shows how this approach may lead to biased transportability estimates in the presence of unmeasured covariates whose distributions differ between the study population and the target population.…”

Improving the impact of HIV pre-exposure prophylaxis implementation in small urban centers among men who have sex with men: An agent-based modelling study

“…Recent research and commentary have attempted to understand the inferences drawn from ABMs within the counterfactual framework in causal inference [ 48 – 51 ]. We refer readers to two recent articles which elucidate the potential bias induced by erroneously assuming the portability of parameter estimates from external populations in the context of ABMs, a concern which may apply to this study [ 50 , 51 ]. We stress the importance of collecting detailed behavioral data in a range of settings, information which is indispensable for parameterizing agent-based models.…”

Using Causal Inference in Field Development Optimization: Application to Unconventional Plays

“…Causal inference is an approach in which the underlying assumed causal structure determines the analysis of the data. Depending on the causal structure, many other methods enable the interpretation of causal effects, such as G-estimation (Murray et al 2017) and the target trial approach (Hernan and Robins 2016). In contrast to estimating a causal effect, the domain of prediction serves the purpose of using available information at a given time to predict an outcome of interest without making inferences about causality, and thus does not require knowledge about the causal structure of the data.…”