(286) Adapting Sufficient Component Cause Models to Pharmacoepidemiology: Mapping to Structural Equation Model Equivalents

Publication Number: 960

Background: Simulations are a critical tool for investigating pharmacoepidemiologic methods questions. Typical simulations do not consider an underlying disease process and often impose simplifying assumptions inconsistent with such a process, leading to unrealistic or even impossible data scenarios. In contrast, data simulated from a sufficient component cause (SCC) model make fewer simplifying assumptions about the underlying causal effect. This gives SCC model-based simulations unique advantages over structural equation model (SEM) based simulations in identifying the limitations of various analytic methods, but fundamental structural differences can make it difficult to translate, compare, and contrast SCC and SEM simulation findings for important pharmacoepidemiologic research questions.

Objectives: Describe and demonstrate a method to map a set of SCC sufficient causes and variable values to SEM coefficients.

Methods: The first step in mapping an SCC simulation to a SEM equivalent is identifying the causes of each relevant variable based on a directed acyclic graph. Next, saturated (or near-saturated) multivariable logistic regression models are fit to predict the probability of each variable conditional on its causes (or, if it has no causes, its overall probability). The coefficients of these logistic regression models are then used as parameters for a “standard” SEM-style simulation, resulting in a study population with similar confounding and overall treatment effect. To test this approach, we simulated a binary treatment X and a binary outcome Y in 500,000 individuals with three confounding variables (two binary, one continuous) using an SCC approach. We then fit saturated multivariable logistic regression models based on the directed acyclic graph created for the SCC model, and simulated another population of 500,000 individuals based on those coefficients. We compared treatment prevalence, outcome rates, and propensity score distributions between the two populations.

Results: Of the 500,000 simulated individuals in the SCC population, 43.5% received X=1. Overall risk of the outcome Y was 27.2%, with a risk of 19.9% in those with X=1 and a risk of 36.6% in those with X=0. In the SEM population of 500,000 simulated individuals, 43.5% received X=1, with an overall outcome risk of 27.4%, a risk of 20.2% in those with X=1 and a risk of 36.7% in those with X=0. Propensity score distributions were virtually identical in both populations.

Conclusions: These findings indicate that a simulated SCC model can be mapped to a unique set of SEM coefficients. This approach will help identify the reasons for any discrepancies in findings from SCC and SEM models and allow in-depth exploration of methods used in pharmacoepidemiologic analyses.