Agent-Based Models for Causal Inference

Abstract

Sound clinical decision making requires evidence-based estimates of the impact of different treatment strategies. In the absence of randomized trials, two potential approaches are agent-based models (ABMs) and the parametric g-formula. Although these methods are mathematically similar, they have generally been considered in isolation. In this dissertation, we bridge the gap between ABMs and the parametric g-formula, in order to improve the use of ABMs for causal inference. In Chapter 1, we describe bias that can occur when ABM inputs or estimates are extrapolated to new populations, and demonstrate the impact of this bias by comparison with the parametric g-formula. We describe the assumptions that are required for extrapolation of an ABM and show that violations of these assumptions produce biased estimates of the risk and causal effect. In Chapter 2, we describe an approach to provide calibration targets for ABMs, and to identify the set of parameters of the ABM that interfere with transportability of the model results to a particular population. We illustrate this approach by comparing the estimates from an existing ABM, the Cost-Effectiveness of Preventing AIDS Complications (CEPAC) model, to estimates from the parametric g-formula applied to a prospective clinical data of HIV-positive individuals under different treatment initiation strategies. In Chapter 3, we focus on the core problem of causal inference from ABMs: how to define and estimate the parameters described in Chapter 2 in light of the bias described in Chapter 1. To illustrate this problem, we consider CEPAC input parameters for opportunistic diseases. We formally define the effect of interest, describe the conditions under which this effect is or is not identifiable, and describe the assumptions required for transportability of this effect. Finally, we show that the estimation of these parameters via a naïve regression analysis approach provides implausible estimates.