Exact Simulation of Longitudinal Data from Marginal Structural Models

Abstract

Simulating longitudinal data from specified marginal structural models is a crucial but challenging task for evaluating causal inference methods and informing study design. While data generation typically proceeds in a fully conditional manner using structural equations according to a temporal ordering, it is difficult to ensure alignment between conditional distributions and the target marginal causal effects, which presents a fundamental challenge. To address this, we propose a flexible and efficient algorithm for simulating longitudinal data that adheres exactly to a specified marginal structural model. Our approach accommodates time-to-event outcomes and extends naturally to survival settings, which are prevalent in applied research. Compared to existing approaches, it offers several advantages: it enables exact simulation from a known causal model rather than relying on approximations; avoids restrictive assumptions about the data-generating process; and remains computationally efficient by requiring only the evaluation of analytical expressions, rather than Monte Carlo methods or numerical integration. Through simulation studies replicating realistic scenarios, we validate the method's accuracy and utility. Our method will facilitate researchers in effectively simulating data with target causal structures for their specific scenarios.